{ "version": "2026-09-07", "description": "LLM / VLM / ASR architecture facts for edge memory sizing: parameters, layers, KV heads, head dimension, maximum context, dtype, vision-tower geometry, published quantised artefact sizes (GGUF) and Jetson benchmark rows, every value with a provenance record (config.json / safetensors index / repository listing = class A; jetson-ai-lab benchmarks = class C). Plus the module memory table and the bytes-per-parameter and runtime-overhead constants the Model Memory Fit engine uses.", "last_updated": "2026-09-07", "confidence": "high", "evidence_class": "A", "documents": { "kv_cache": { "publisher": "NVIDIA", "document": "Mastering LLM Techniques: Inference Optimization (NVIDIA Technical Blog)", "doc_id": "nvidia-blog:mastering-llm-techniques-inference-optimization", "section": "Key-value caching", "url": "https://developer.nvidia.com/blog/mastering-llm-techniques-inference-optimization/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "KV cache size per token = 2 × (num_layers) × (num_heads × dim_head) × precision_in_bytes" }, "gguf_q8_0": { "publisher": "ggml-org (llama.cpp)", "document": "ggml quantization block formats (ggml-common.h)", "doc_id": "github:ggml-org/llama.cpp/ggml-common.h", "section": "block_q8_0", "url": "https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-common.h", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "QK8_0 32 … block_q8_0 { ggml_half d; int8_t qs[QK8_0]; }" }, "jetson_unified_memory": { "publisher": "NVIDIA", "document": "CUDA for Tegra — Memory Management", "doc_id": "nvidia-docs:cuda-for-tegra-appnote", "section": "Memory Management", "url": "https://docs.nvidia.com/cuda/cuda-for-tegra-appnote/index.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "the iGPU and the CPU share the same SoC DRAM" } }, "quants": { "fp16": { "label": "FP16 / BF16", "fixed": 2, "class": "A", "note": "16-bit storage width", "derived_bytes_per_param": 2 }, "fp8": { "label": "FP8 (E4M3)", "fixed": 1, "class": "A", "note": "8-bit storage width; TensorRT-LLM / vLLM on Blackwell (Thor)", "derived_bytes_per_param": 1 }, "int8": { "label": "INT8 / GGUF Q8_0", "nominal": 1.0625, "class": "D", "note": "Q8_0: 32 int8 weights + one fp16 scale per block = 8.5 bits per weight (DOC.gguf_q8_0)", "derived_bytes_per_param": 1.064 }, "q4": { "label": "GGUF Q4_K_M / INT4 (AWQ, MLC q4f16)", "nominal": 0.625, "class": "D", "note": "K-quant mixes 4- and 6-bit blocks with fp16 scales; registry median ratio applies", "derived_bytes_per_param": 0.625 }, "nvfp4": { "label": "NVFP4", "nominal": 0.5625, "class": "D", "note": "4-bit weights + one fp8 scale per 16-weight block = 4.5 bits per weight", "derived_bytes_per_param": 0.5625 } }, "runtime_overhead": { "llama_cpp": { "label": "llama.cpp / Ollama", "base_mb": 400, "workspace_pct": 6, "kv_bytes": 2, "note": "CUDA context + compute buffer for the prompt batch (n_batch 512)" }, "mlc": { "label": "MLC LLM", "base_mb": 400, "workspace_pct": 6, "kv_bytes": 2, "note": "TVM runtime + CUDA context; q4f16 KV in fp16" }, "tensorrt_llm": { "label": "TensorRT-LLM", "base_mb": 600, "workspace_pct": 10, "kv_bytes": 2, "note": "engine activations + CUDA context; paged KV pool sized to max context" }, "vllm": { "label": "vLLM", "base_mb": 800, "workspace_pct": 10, "kv_bytes": 2, "note": "CUDA graphs + activations; vLLM pre-allocates gpu_memory_utilization × total, so the fit here is the minimum" }, "transformers": { "label": "Hugging Face Transformers (PyTorch)", "base_mb": 900, "workspace_pct": 25, "kv_bytes": 2, "note": "PyTorch allocator + eager activations; no paged KV" } }, "schema": { "models..architecture": "{ layers, hidden, heads, kv_heads, head_dim, max_ctx, dtype, vocab, head_dim_derived, sources[] }", "models..quantized_artefacts[]": "{ quant (fp16|int8|q4|fp8|nvfp4), format, file, size_gb, sources[] }", "models..jetson_measurements[]": "{ module, runtime, quant, memory_gb, tokens_per_s, sources[] }", "models..image_tokens": "VLM tokens added to the context per image with the derivation basis and evidence class", "module_memory.": "{ memory_gb, sources[] } — NVIDIA module memory (class A)", "formula": "kv_cache_bytes = 2 × layers × kv_heads × head_dim × context_tokens × bytes_per_element × sequences" }, "module_memory": { "jetson_orin_nano": { "memory_gb": 8, "sources": [ { "publisher": "NVIDIA", "document": "Jetson Orin Nano / NX / AGX Orin technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "8GB 128-bit LPDDR5| 102 GB/s" } ] }, "jetson_orin_nx_8gb": { "memory_gb": 8, "sources": [ { "publisher": "NVIDIA", "document": "Jetson Orin Nano / NX / AGX Orin technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "8GB 128-bit LPDDR5| 102.4GB/s" } ] }, "jetson_orin_nx": { "memory_gb": 16, "sources": [ { "publisher": "NVIDIA", "document": "Jetson Orin Nano / NX / AGX Orin technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "16GB 128-bit LPDDR5| 102.4GB/s" } ] }, "jetson_agx_orin_32gb": { "memory_gb": 32, "sources": [ { "publisher": "NVIDIA", "document": "Jetson Orin Nano / NX / AGX Orin technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "32GB 256-bit LPDDR5| 204.8GB/s" } ] }, "jetson_agx_orin": { "memory_gb": 64, "sources": [ { "publisher": "NVIDIA", "document": "Jetson Orin Nano / NX / AGX Orin technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "64GB 256-bit LPDDR5 | 204.8GB/s" } ] }, "jetson_thor_t5000": { "memory_gb": 128, "sources": [ { "publisher": "NVIDIA", "document": "Jetson AGX Thor / Jetson T5000 / T4000 technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "128 GB 256-bit LPDDR5X| 273 GB/s" } ] }, "jetson_thor_t4000": { "memory_gb": 64, "sources": [ { "publisher": "NVIDIA", "document": "Jetson AGX Thor / Jetson T5000 / T4000 technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "64 GB 256-bit LPDDR5X| 273 GB/s" } ] }, "jetson_orin_nano_super": { "memory_gb": 8, "sources": [ { "publisher": "NVIDIA", "document": "Jetson Orin Nano / NX / AGX Orin technical specifications table", "doc_id": "nvidia.com:jetson-modules-spec-table", "url": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "8GB 128-bit LPDDR5| 102 GB/s" } ], "alias_of": "jetson_orin_nano" } }, "models": { "llama-3.2-1b-instruct": { "label": "Llama 3.2 1B Instruct", "family": "llama", "type": "llm", "hf_repo": "meta-llama/Llama-3.2-1B-Instruct", "params_b": 1.24, "text_params_b": 1.24, "params_sources": [ { "publisher": "Hugging Face (mirror: unsloth/Llama-3.2-1B-Instruct, identical weights to meta-llama repo)", "document": "model.safetensors (blob metadata via HF API)", "doc_id": "hf:unsloth/Llama-3.2-1B-Instruct", "url": "https://huggingface.co/api/models/unsloth/Llama-3.2-1B-Instruct?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "model.safetensors size 2471645608 bytes" } ], "architecture": { "layers": 16, "hidden": 2048, "heads": 32, "kv_heads": 8, "head_dim": 64, "max_ctx": 131072, "dtype": "bfloat16", "vocab": 128256, "head_dim_derived": false, "sources": [ { "publisher": "Hugging Face (mirror: unsloth/Llama-3.2-1B-Instruct)", "document": "config.json", "doc_id": "hf:unsloth/Llama-3.2-1B-Instruct", "url": "https://huggingface.co/unsloth/Llama-3.2-1B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_hidden_layers\": 16, \"hidden_size\": 2048, \"head_dim\": 64" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Llama-3.2-1B-Instruct-Q4_K_M.gguf", "size_gb": 0.808, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Llama-3.2-1B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Llama-3.2-1B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-1B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Llama-3.2-1B-Instruct-Q4_K_M.gguf size 807694464 bytes" } ] }, { "quant": "int8", "format": "GGUF Q8_0", "file": "Llama-3.2-1B-Instruct-Q8_0.gguf", "size_gb": 1.321, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Llama-3.2-1B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Llama-3.2-1B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-1B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Llama-3.2-1B-Instruct-Q8_0.gguf size 1321083008 bytes" } ] } ], "jetson_measurements": [ { "module": "jetson_orin_nano", "runtime": "MLC", "quant": "INT4 (q4f16_ft)", "memory_gb": null, "tokens_per_s": 54.8, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "tutorial_slm.html", "doc_id": "jetson-ai-lab:tutorial_slm.html", "url": "https://www.jetson-ai-lab.com/archive/tutorial_slm.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Llama-3.2-1B Jetson Orin Nano 54.8 tokens/sec, Jetson AGX Orin 163.9 tokens/sec" } ] }, { "module": "jetson_agx_orin", "runtime": "MLC", "quant": "INT4 (q4f16_ft)", "memory_gb": null, "tokens_per_s": 163.9, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "tutorial_slm.html", "doc_id": "jetson-ai-lab:tutorial_slm.html", "url": "https://www.jetson-ai-lab.com/archive/tutorial_slm.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Llama-3.2-1B Jetson Orin Nano 54.8 tokens/sec, Jetson AGX Orin 163.9 tokens/sec" } ] } ], "notes": "params_b computed as safetensors total_size / 2 bytes (bf16); meta-llama org repo returns 401 (gated) so unsloth mirror (byte-identical weights) used as source for config.json and file size." }, "llama-3.2-3b-instruct": { "label": "Llama 3.2 3B Instruct", "family": "llama", "type": "llm", "hf_repo": "meta-llama/Llama-3.2-3B-Instruct", "params_b": 3.213, "text_params_b": 3.213, "params_sources": [ { "publisher": "Hugging Face (mirror: unsloth/Llama-3.2-3B-Instruct)", "document": "model.safetensors.index.json", "doc_id": "hf:unsloth/Llama-3.2-3B-Instruct", "url": "https://huggingface.co/unsloth/Llama-3.2-3B-Instruct/raw/main/model.safetensors.index.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "\"total_size\": 6425499648" } ], "architecture": { "layers": 28, "hidden": 3072, "heads": 24, "kv_heads": 8, "head_dim": 128, "max_ctx": 131072, "dtype": "bfloat16", "vocab": 128256, "head_dim_derived": false, "sources": [ { "publisher": "Hugging Face (mirror: unsloth/Llama-3.2-3B-Instruct)", "document": "config.json", "doc_id": "hf:unsloth/Llama-3.2-3B-Instruct", "url": "https://huggingface.co/unsloth/Llama-3.2-3B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_hidden_layers\": 28, \"hidden_size\": 3072, \"head_dim\": 128" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Llama-3.2-3B-Instruct-Q4_K_M.gguf", "size_gb": 2.019, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Llama-3.2-3B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Llama-3.2-3B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-3B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Llama-3.2-3B-Instruct-Q4_K_M.gguf size 2019377696 bytes" } ] }, { "quant": "int8", "format": "GGUF Q8_0", "file": "Llama-3.2-3B-Instruct-Q8_0.gguf", "size_gb": 3.422, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Llama-3.2-3B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Llama-3.2-3B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Llama-3.2-3B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Llama-3.2-3B-Instruct-Q8_0.gguf size 3421899296 bytes" } ] } ], "jetson_measurements": [ { "module": "jetson_orin_nano", "runtime": "MLC", "quant": "INT4 (q4f16_ft)", "memory_gb": null, "tokens_per_s": 27.7, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "tutorial_slm.html", "doc_id": "jetson-ai-lab:tutorial_slm.html", "url": "https://www.jetson-ai-lab.com/archive/tutorial_slm.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Llama-3.2-3B Jetson Orin Nano 27.7 tokens/sec, Jetson AGX Orin 80.4 tokens/sec" } ] }, { "module": "jetson_agx_orin", "runtime": "MLC", "quant": "INT4 (q4f16_ft)", "memory_gb": null, "tokens_per_s": 80.4, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "tutorial_slm.html", "doc_id": "jetson-ai-lab:tutorial_slm.html", "url": "https://www.jetson-ai-lab.com/archive/tutorial_slm.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Llama-3.2-3B Jetson Orin Nano 27.7 tokens/sec, Jetson AGX Orin 80.4 tokens/sec" } ] }, { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark)", "quant": null, "memory_gb": null, "tokens_per_s": 43.07, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Llama 3.2 3B: Jetson Orin Nano 27.7, Jetson Orin Nano Super 43.07 (perf gain 1.55x)" } ] } ], "notes": "meta-llama org repo returns 401 (gated); unsloth mirror (byte-identical weights) used." }, "llama-3.1-8b-instruct": { "label": "Llama 3.1 8B Instruct", "family": "llama", "type": "llm", "hf_repo": "meta-llama/Meta-Llama-3.1-8B-Instruct", "params_b": 8.03, "text_params_b": 8.03, "params_sources": [ { "publisher": "Hugging Face (mirror: NousResearch/Meta-Llama-3.1-8B-Instruct)", "document": "model.safetensors.index.json", "doc_id": "hf:NousResearch/Meta-Llama-3.1-8B-Instruct", "url": "https://huggingface.co/NousResearch/Meta-Llama-3.1-8B-Instruct/raw/main/model.safetensors.index.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "\"total_size\": 16060522496" } ], "architecture": { "layers": 32, "hidden": 4096, "heads": 32, "kv_heads": 8, "head_dim": 128, "max_ctx": 131072, "dtype": "bfloat16", "vocab": 128256, "head_dim_derived": false, "sources": [ { "publisher": "Hugging Face (mirror: NousResearch/Meta-Llama-3.1-8B-Instruct)", "document": "config.json", "doc_id": "hf:NousResearch/Meta-Llama-3.1-8B-Instruct", "url": "https://huggingface.co/NousResearch/Meta-Llama-3.1-8B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_hidden_layers\": 32, \"hidden_size\": 4096, \"num_key_value_heads\": 8" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf", "size_gb": 4.921, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Meta-Llama-3.1-8B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Meta-Llama-3.1-8B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf size 4920739232 bytes" } ] }, { "quant": "int8", "format": "GGUF Q8_0", "file": "Meta-Llama-3.1-8B-Instruct-Q8_0.gguf", "size_gb": 8.541, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Meta-Llama-3.1-8B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Meta-Llama-3.1-8B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Meta-Llama-3.1-8B-Instruct-Q8_0.gguf size 8540775840 bytes" } ] } ], "jetson_measurements": [ { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark)", "quant": null, "memory_gb": null, "tokens_per_s": 19.14, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Llama 3.1 8B: Jetson Orin Nano 14, Jetson Orin Nano Super 19.14 (perf gain 1.37x)" } ] } ], "notes": "meta-llama org repo returns 401 (gated); NousResearch mirror (byte-identical weights) used. HF card 'Params' table lists 8B." }, "qwen2.5-1.5b-instruct": { "label": "Qwen2.5 1.5B Instruct", "family": "qwen2.5", "type": "llm", "hf_repo": "Qwen/Qwen2.5-1.5B-Instruct", "params_b": 1.54, "text_params_b": 1.54, "params_sources": [ { "publisher": "Qwen (Alibaba)", "document": "Qwen/Qwen2.5-1.5B-Instruct README.md", "doc_id": "hf:Qwen/Qwen2.5-1.5B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Number of Parameters: 1.54B" } ], "architecture": { "layers": 28, "hidden": 1536, "heads": 12, "kv_heads": 2, "head_dim": 128, "max_ctx": 32768, "dtype": "bfloat16", "vocab": 151936, "head_dim_derived": false, "sources": [ { "publisher": "Qwen (Alibaba)", "document": "config.json", "doc_id": "hf:Qwen/Qwen2.5-1.5B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"hidden_size\": 1536, \"num_attention_heads\": 12, \"num_key_value_heads\": 2" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Qwen2.5-1.5B-Instruct-Q4_K_M.gguf", "size_gb": 0.986, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Qwen2.5-1.5B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Qwen2.5-1.5B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Qwen2.5-1.5B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Qwen2.5-1.5B-Instruct-Q4_K_M.gguf size 986048768 bytes" } ] } ], "jetson_measurements": [], "notes": "head_dim not in config.json; derived as hidden_size/num_attention_heads=1536/12=128." }, "qwen2.5-3b-instruct": { "label": "Qwen2.5 3B Instruct", "family": "qwen2.5", "type": "llm", "hf_repo": "Qwen/Qwen2.5-3B-Instruct", "params_b": 3.09, "text_params_b": 3.09, "params_sources": [ { "publisher": "Qwen (Alibaba)", "document": "Qwen/Qwen2.5-3B-Instruct README.md", "doc_id": "hf:Qwen/Qwen2.5-3B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Number of Parameters: 3.09B" } ], "architecture": { "layers": 36, "hidden": 2048, "heads": 16, "kv_heads": 2, "head_dim": 128, "max_ctx": 32768, "dtype": "bfloat16", "vocab": 151936, "head_dim_derived": false, "sources": [ { "publisher": "Qwen (Alibaba)", "document": "config.json", "doc_id": "hf:Qwen/Qwen2.5-3B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"hidden_size\": 2048, \"num_attention_heads\": 16, \"num_key_value_heads\": 2" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Qwen2.5-3B-Instruct-Q4_K_M.gguf", "size_gb": 1.93, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Qwen2.5-3B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Qwen2.5-3B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Qwen2.5-3B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Qwen2.5-3B-Instruct-Q4_K_M.gguf size 1929903264 bytes" } ] } ], "jetson_measurements": [], "notes": "head_dim derived as hidden_size/num_attention_heads=2048/16=128." }, "qwen2.5-7b-instruct": { "label": "Qwen2.5 7B Instruct", "family": "qwen2.5", "type": "llm", "hf_repo": "Qwen/Qwen2.5-7B-Instruct", "params_b": 7.61, "text_params_b": 7.61, "params_sources": [ { "publisher": "Qwen (Alibaba)", "document": "Qwen/Qwen2.5-7B-Instruct README.md", "doc_id": "hf:Qwen/Qwen2.5-7B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Number of Parameters: 7.61B" } ], "architecture": { "layers": 28, "hidden": 3584, "heads": 28, "kv_heads": 4, "head_dim": 128, "max_ctx": 32768, "dtype": "bfloat16", "vocab": 152064, "head_dim_derived": false, "sources": [ { "publisher": "Qwen (Alibaba)", "document": "config.json", "doc_id": "hf:Qwen/Qwen2.5-7B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"hidden_size\": 3584, \"num_attention_heads\": 28, \"num_key_value_heads\": 4" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Qwen2.5-7B-Instruct-Q4_K_M.gguf", "size_gb": 4.683, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Qwen2.5-7B-Instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Qwen2.5-7B-Instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Qwen2.5-7B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Qwen2.5-7B-Instruct-Q4_K_M.gguf size 4683074240 bytes" } ] } ], "jetson_measurements": [ { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark)", "quant": null, "memory_gb": null, "tokens_per_s": 21.75, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Qwen2.5 7B: Jetson Orin Nano 14.2, Jetson Orin Nano Super 21.75 (perf gain 1.53x)" } ] } ], "notes": "head_dim derived as hidden_size/num_attention_heads=3584/28=128." }, "qwen2.5-vl-3b-instruct": { "label": "Qwen2.5-VL 3B Instruct", "family": "qwen2.5-vl", "type": "vlm", "hf_repo": "Qwen/Qwen2.5-VL-3B-Instruct", "params_b": 3.755, "text_params_b": 3.126, "params_sources": [ { "publisher": "Hugging Face", "document": "model.safetensors.index.json", "doc_id": "hf:Qwen/Qwen2.5-VL-3B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/main/model.safetensors.index.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "\"total_size\": 7509245952" } ], "architecture": { "layers": 36, "hidden": 2048, "heads": 16, "kv_heads": 2, "head_dim": 128, "max_ctx": 128000, "dtype": "bfloat16", "vocab": 151936, "head_dim_derived": false, "sources": [ { "publisher": "Qwen (Alibaba)", "document": "config.json", "doc_id": "hf:Qwen/Qwen2.5-VL-3B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"hidden_size\": 2048, \"num_attention_heads\": 16, \"num_key_value_heads\": 2, \"max_position_embeddings\": 128000" } ] }, "vision": { "layers": 32, "hidden": 1280, "patch_size": 14, "image_size": null, "model_type": null, "sources": [ { "publisher": "Qwen (Alibaba)", "document": "config.json vision_config", "doc_id": "hf:Qwen/Qwen2.5-VL-3B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"depth\": 32, \"hidden_size\": 1280, \"num_heads\": 16, \"patch_size\": 14" } ], "mmproj": { "format": "GGUF mmproj Q8_0 (vision tower)", "file": "mmproj-Qwen2.5-VL-3B-Instruct-Q8_0.gguf", "size_gb": 0.845, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "Qwen2.5-VL-3B-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/Qwen2.5-VL-3B-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/Qwen2.5-VL-3B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "mmproj-Qwen2.5-VL-3B-Instruct-Q8_0.gguf size 844757728 bytes" } ] } }, "image_tokens": { "tokens": 1280, "class": "D", "basis": "dynamic: (H/28)×(W/28) tokens after 2×2 spatial merge; 1280 assumed for a ~1 MP image" }, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M (text)", "file": "Qwen2.5-VL-3B-Instruct-Q4_K_M.gguf", "size_gb": 1.93, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "Qwen2.5-VL-3B-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/Qwen2.5-VL-3B-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/Qwen2.5-VL-3B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Qwen2.5-VL-3B-Instruct-Q4_K_M.gguf size 1929901056 bytes" } ] } ], "jetson_measurements": [], "notes": "params_b (3.75B) is the FULL model (text backbone + vision encoder) from safetensors total_size; the '3B' name refers to the LLM backbone only. Vision GGUF requires a separate mmproj file loaded alongside the text GGUF in llama.cpp." }, "qwen2.5-vl-7b-instruct": { "label": "Qwen2.5-VL 7B Instruct", "family": "qwen2.5-vl", "type": "vlm", "hf_repo": "Qwen/Qwen2.5-VL-7B-Instruct", "params_b": 8.292, "text_params_b": 7.663, "params_sources": [ { "publisher": "Hugging Face", "document": "model.safetensors.index.json", "doc_id": "hf:Qwen/Qwen2.5-VL-7B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/main/model.safetensors.index.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "\"total_size\": 16584333312" } ], "architecture": { "layers": 28, "hidden": 3584, "heads": 28, "kv_heads": 4, "head_dim": 128, "max_ctx": 128000, "dtype": "bfloat16", "vocab": 152064, "head_dim_derived": false, "sources": [ { "publisher": "Qwen (Alibaba)", "document": "config.json", "doc_id": "hf:Qwen/Qwen2.5-VL-7B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"hidden_size\": 3584, \"num_attention_heads\": 28, \"num_key_value_heads\": 4, \"max_position_embeddings\": 128000" } ] }, "vision": { "layers": 32, "hidden": 1280, "patch_size": 14, "image_size": null, "model_type": null, "sources": [ { "publisher": "Qwen (Alibaba)", "document": "config.json vision_config", "doc_id": "hf:Qwen/Qwen2.5-VL-7B-Instruct", "url": "https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"depth\": 32, \"hidden_size\": 1280, \"num_heads\": 16, \"patch_size\": 14" } ], "mmproj": { "format": "GGUF mmproj Q8_0 (vision tower)", "file": "mmproj-Qwen2.5-VL-7B-Instruct-Q8_0.gguf", "size_gb": 0.853, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "Qwen2.5-VL-7B-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/Qwen2.5-VL-7B-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/Qwen2.5-VL-7B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "mmproj-Qwen2.5-VL-7B-Instruct-Q8_0.gguf size 853119712 bytes" } ] } }, "image_tokens": { "tokens": 1280, "class": "D", "basis": "dynamic: (H/28)×(W/28) tokens after 2×2 spatial merge; 1280 assumed for a ~1 MP image" }, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M (text)", "file": "Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf", "size_gb": 4.683, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "Qwen2.5-VL-7B-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/Qwen2.5-VL-7B-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/Qwen2.5-VL-7B-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf size 4683072032 bytes" } ] } ], "jetson_measurements": [], "notes": "params_b (8.29B) is the FULL model (text backbone + vision encoder) from safetensors total_size; the '7B' name refers to the LLM backbone only." }, "phi-3.5-mini-instruct": { "label": "Phi-3.5-mini Instruct", "family": "phi3", "type": "llm", "hf_repo": "microsoft/Phi-3.5-mini-instruct", "params_b": 3.8, "text_params_b": 3.8, "params_sources": [ { "publisher": "Microsoft", "document": "Phi-3.5-mini-instruct README.md", "doc_id": "hf:microsoft/Phi-3.5-mini-instruct", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Phi-3.5-mini has 3.8B parameters and is a dense decoder-only Transformer model" } ], "architecture": { "layers": 32, "hidden": 3072, "heads": 32, "kv_heads": 32, "head_dim": 96, "max_ctx": 131072, "dtype": "bfloat16", "vocab": 32064, "head_dim_derived": false, "sources": [ { "publisher": "Microsoft", "document": "config.json", "doc_id": "hf:microsoft/Phi-3.5-mini-instruct", "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 32, \"num_key_value_heads\": 32, \"hidden_size\": 3072" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Phi-3.5-mini-instruct-Q4_K_M.gguf", "size_gb": 2.393, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Phi-3.5-mini-instruct-GGUF repo file listing", "doc_id": "hf:bartowski/Phi-3.5-mini-instruct-GGUF", "url": "https://huggingface.co/api/models/bartowski/Phi-3.5-mini-instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Phi-3.5-mini-instruct-Q4_K_M.gguf size 2393232672 bytes" } ] } ], "jetson_measurements": [], "notes": "no explicit head_dim field in config.json; derived as hidden_size/num_attention_heads=3072/32=96 (no GQA, num_key_value_heads==num_attention_heads)." }, "phi-4-mini-instruct": { "label": "Phi-4-mini Instruct", "family": "phi3(phi4)", "type": "llm", "hf_repo": "microsoft/Phi-4-mini-instruct", "params_b": 3.8, "text_params_b": 3.8, "params_sources": [ { "publisher": "Microsoft", "document": "Phi-4-mini-instruct README.md", "doc_id": "hf:microsoft/Phi-4-mini-instruct", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "the 3.8B parameters Phi-4-mini-instruct model was compared with a set of models" } ], "architecture": { "layers": 32, "hidden": 3072, "heads": 24, "kv_heads": 8, "head_dim": 128, "max_ctx": 131072, "dtype": "bfloat16", "vocab": 200064, "head_dim_derived": true, "sources": [ { "publisher": "Microsoft", "document": "config.json", "doc_id": "hf:microsoft/Phi-4-mini-instruct", "url": "https://huggingface.co/microsoft/Phi-4-mini-instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 24, \"num_key_value_heads\": 8, \"partial_rotary_factor\": 0.75" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Phi-4-mini-instruct-Q4_K_M.gguf", "size_gb": 2.492, "sources": [ { "publisher": "unsloth (community GGUF)", "document": "Phi-4-mini-instruct-GGUF repo file listing", "doc_id": "hf:unsloth/Phi-4-mini-instruct-GGUF", "url": "https://huggingface.co/api/models/unsloth/Phi-4-mini-instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Phi-4-mini-instruct-Q4_K_M.gguf size 2491874272 bytes" } ] } ], "jetson_measurements": [], "notes": "head_dim left null: hidden_size/num_attention_heads=3072/24=128 nominal, but config sets partial_rotary_factor=0.75 (rotary applied to 96 of 128 dims) -- could not find an explicit head_dim field to quote, so not stated as a hard number. bartowski/Phi-4-mini-instruct-GGUF returned 401 (repo not found under that name); unsloth/Phi-4-mini-instruct-GGUF used instead." }, "gemma-2-2b-it": { "label": "Gemma 2 2B IT", "family": "gemma2", "type": "llm", "hf_repo": "google/gemma-2-2b-it", "params_b": 2.614, "text_params_b": 2.614, "params_sources": [ { "publisher": "Hugging Face (mirror: unsloth/gemma-2-2b-it, identical weights to google/gemma-2-2b-it)", "document": "model.safetensors (blob metadata via HF API)", "doc_id": "hf:unsloth/gemma-2-2b-it", "url": "https://huggingface.co/api/models/unsloth/gemma-2-2b-it?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "model.safetensors size 5228717512 bytes" } ], "architecture": { "layers": 26, "hidden": 2304, "heads": 8, "kv_heads": 4, "head_dim": 256, "max_ctx": 8192, "dtype": "bfloat16", "vocab": 256000, "head_dim_derived": false, "sources": [ { "publisher": "Hugging Face (mirror: unsloth/gemma-2-2b-it)", "document": "config.json", "doc_id": "hf:unsloth/gemma-2-2b-it", "url": "https://huggingface.co/unsloth/gemma-2-2b-it/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 8, \"num_key_value_heads\": 4, \"head_dim\": 256" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "gemma-2-2b-it-Q4_K_M.gguf", "size_gb": 1.709, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "gemma-2-2b-it-GGUF repo file listing", "doc_id": "hf:bartowski/gemma-2-2b-it-GGUF", "url": "https://huggingface.co/api/models/bartowski/gemma-2-2b-it-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "gemma-2-2b-it-Q4_K_M.gguf size 1708582752 bytes" } ] } ], "jetson_measurements": [ { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark)", "quant": null, "memory_gb": null, "tokens_per_s": 34.97, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "Gemma 2 2B: Jetson Orin Nano 21.5, Jetson Orin Nano Super 34.97 (perf gain 1.63x)" } ] } ], "notes": "google/gemma-2-2b-it config.json/README return 401 (gated); unsloth mirror (byte-identical weights) used." }, "gemma-3-4b-it": { "label": "Gemma 3 4B IT", "family": "gemma3", "type": "vlm", "hf_repo": "google/gemma-3-4b-it", "params_b": 4.3, "text_params_b": 3.87, "params_sources": [ { "publisher": "Hugging Face (mirror: unsloth/gemma-3-4b-it, identical weights to google/gemma-3-4b-it)", "document": "model.safetensors.index.json", "doc_id": "hf:unsloth/gemma-3-4b-it", "url": "https://huggingface.co/unsloth/gemma-3-4b-it/raw/main/model.safetensors.index.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "\"total_size\": 8600158944" } ], "architecture": { "layers": 34, "hidden": 2560, "heads": 8, "kv_heads": 4, "head_dim": 256, "max_ctx": 131072, "dtype": "bfloat16", "vocab": 262208, "head_dim_derived": false, "sources": [ { "publisher": "Hugging Face (mirror: unsloth/gemma-3-4b-it)", "document": "config.json text_config", "doc_id": "hf:unsloth/gemma-3-4b-it", "url": "https://huggingface.co/unsloth/gemma-3-4b-it/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 8, \"num_key_value_heads\": 4, \"head_dim\": 256" } ] }, "vision": { "layers": 27, "hidden": 1152, "patch_size": 14, "image_size": 896, "model_type": "siglip_vision_model", "sources": [ { "publisher": "Hugging Face (mirror: unsloth/gemma-3-4b-it)", "document": "config.json vision_config", "doc_id": "hf:unsloth/gemma-3-4b-it", "url": "https://huggingface.co/unsloth/gemma-3-4b-it/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"model_type\": \"siglip_vision_model\", \"hidden_size\": 1152, \"image_size\": 896" } ] }, "image_tokens": { "tokens": 256, "class": "D", "basis": "SigLIP 896 / 14 = 64² patches pooled to 256 soft tokens per image (Gemma 3 technical report §2)" }, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "gemma-3-4b-it-Q4_K_M.gguf", "size_gb": 2.49, "sources": [ { "publisher": "unsloth (community GGUF)", "document": "gemma-3-4b-it-GGUF repo file listing", "doc_id": "hf:unsloth/gemma-3-4b-it-GGUF", "url": "https://huggingface.co/api/models/unsloth/gemma-3-4b-it-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "gemma-3-4b-it-Q4_K_M.gguf size 2489894016 bytes" } ] } ], "jetson_measurements": [], "notes": "google/gemma-3-4b-it config.json returns 401 (gated); unsloth mirror (byte-identical weights) used. bartowski/gemma-3-4b-it-GGUF returned 401 (repo not found under that name); unsloth/gemma-3-4b-it-GGUF used instead. params_b (4.3B) is the FULL model (text + SigLIP vision encoder)." }, "smolvlm-256m-instruct": { "label": "SmolVLM 256M Instruct", "family": "smolvlm", "type": "vlm", "hf_repo": "HuggingFaceTB/SmolVLM-256M-Instruct", "params_b": 0.2565, "text_params_b": 0.172, "params_sources": [ { "publisher": "Hugging Face", "document": "model.safetensors (blob metadata via HF API)", "doc_id": "hf:HuggingFaceTB/SmolVLM-256M-Instruct", "url": "https://huggingface.co/api/models/HuggingFaceTB/SmolVLM-256M-Instruct?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "model.safetensors size 513028808 bytes" } ], "architecture": { "layers": 30, "hidden": 576, "heads": 9, "kv_heads": 3, "head_dim": 64, "max_ctx": 8192, "dtype": "bfloat16", "vocab": 49280, "head_dim_derived": false, "sources": [ { "publisher": "HuggingFaceTB", "document": "config.json text_config", "doc_id": "hf:HuggingFaceTB/SmolVLM-256M-Instruct", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-256M-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 9, \"num_key_value_heads\": 3, \"head_dim\": 64" } ] }, "vision": { "layers": 12, "hidden": 768, "patch_size": 16, "image_size": 512, "model_type": "idefics3 (SigLIP-based, 93M-param encoder per model card)", "sources": [ { "publisher": "HuggingFaceTB", "document": "config.json vision_config", "doc_id": "hf:HuggingFaceTB/SmolVLM-256M-Instruct", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-256M-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"hidden_size\": 768, \"num_hidden_layers\": 12, \"image_size\": 512" } ], "mmproj": { "format": "GGUF mmproj Q8_0 (vision)", "file": "mmproj-SmolVLM-256M-Instruct-Q8_0.gguf", "size_gb": 0.104, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "SmolVLM-256M-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/SmolVLM-256M-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-256M-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "mmproj-SmolVLM-256M-Instruct-Q8_0.gguf size 103769856 bytes" } ] } }, "image_tokens": { "tokens": 64, "class": "D", "basis": "512² patch / 16 = 32² tokens, pixel-shuffle ×4 → 64 tokens per 512×512 tile; one tile assumed" }, "quantized_artefacts": [ { "quant": "int8", "format": "GGUF Q8_0 (text)", "file": "SmolVLM-256M-Instruct-Q8_0.gguf", "size_gb": 0.175, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "SmolVLM-256M-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/SmolVLM-256M-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-256M-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "SmolVLM-256M-Instruct-Q8_0.gguf size 175054528 bytes" } ] } ], "jetson_measurements": [], "notes": "No Q4_K_M published for this size; Q8_0 is the smallest quant in the ggml-org repo. Model card states 'run inference on one image with under 1GB of GPU RAM' (generic GPU, not Jetson-specific)." }, "smolvlm-500m-instruct": { "label": "SmolVLM 500M Instruct", "family": "smolvlm", "type": "vlm", "hf_repo": "HuggingFaceTB/SmolVLM-500M-Instruct", "params_b": 0.5075, "text_params_b": 0.422, "params_sources": [ { "publisher": "Hugging Face", "document": "model.safetensors (blob metadata via HF API)", "doc_id": "hf:HuggingFaceTB/SmolVLM-500M-Instruct", "url": "https://huggingface.co/api/models/HuggingFaceTB/SmolVLM-500M-Instruct?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "model.safetensors size 1015025832 bytes" } ], "architecture": { "layers": 32, "hidden": 960, "heads": 15, "kv_heads": 5, "head_dim": 64, "max_ctx": 8192, "dtype": "bfloat16", "vocab": 49280, "head_dim_derived": false, "sources": [ { "publisher": "HuggingFaceTB", "document": "config.json text_config", "doc_id": "hf:HuggingFaceTB/SmolVLM-500M-Instruct", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 15, \"num_key_value_heads\": 5, \"head_dim\": 64" } ] }, "vision": { "layers": 12, "hidden": 768, "patch_size": 16, "image_size": 512, "model_type": "idefics3 (SigLIP-based)", "sources": [ { "publisher": "HuggingFaceTB", "document": "config.json vision_config", "doc_id": "hf:HuggingFaceTB/SmolVLM-500M-Instruct", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"hidden_size\": 768, \"num_hidden_layers\": 12, \"image_size\": 512" } ], "mmproj": { "format": "GGUF mmproj Q8_0 (vision)", "file": "mmproj-SmolVLM-500M-Instruct-Q8_0.gguf", "size_gb": 0.109, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "SmolVLM-500M-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/SmolVLM-500M-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-500M-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "mmproj-SmolVLM-500M-Instruct-Q8_0.gguf size 108783360 bytes" } ] } }, "image_tokens": { "tokens": 64, "class": "D", "basis": "512² patch / 16 = 32² tokens, pixel-shuffle ×4 → 64 tokens per 512×512 tile; one tile assumed" }, "quantized_artefacts": [ { "quant": "int8", "format": "GGUF Q8_0 (text)", "file": "SmolVLM-500M-Instruct-Q8_0.gguf", "size_gb": 0.437, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "SmolVLM-500M-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/SmolVLM-500M-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-500M-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "SmolVLM-500M-Instruct-Q8_0.gguf size 436806912 bytes" } ] } ], "jetson_measurements": [], "notes": "Model card states 'run inference on one image with 1.23GB of GPU RAM' (generic GPU, not Jetson-specific)." }, "smolvlm-2.2b-instruct": { "label": "SmolVLM 2.2B Instruct", "family": "smolvlm", "type": "vlm", "hf_repo": "HuggingFaceTB/SmolVLM-Instruct", "params_b": 2.246, "text_params_b": 1.816, "params_sources": [ { "publisher": "Hugging Face", "document": "model.safetensors (blob metadata via HF API)", "doc_id": "hf:HuggingFaceTB/SmolVLM-Instruct", "url": "https://huggingface.co/api/models/HuggingFaceTB/SmolVLM-Instruct?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "model.safetensors size 4492630912 bytes" } ], "architecture": { "layers": 24, "hidden": 2048, "heads": 32, "kv_heads": 32, "head_dim": 64, "max_ctx": 16384, "dtype": "bfloat16", "vocab": 49155, "head_dim_derived": false, "sources": [ { "publisher": "HuggingFaceTB", "document": "config.json text_config", "doc_id": "hf:HuggingFaceTB/SmolVLM-Instruct", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 32, \"num_key_value_heads\": 32, \"head_dim\": 64" } ] }, "vision": { "layers": 27, "hidden": 1152, "patch_size": 14, "image_size": 384, "model_type": "idefics3 (SigLIP-SO400M vision encoder)", "sources": [ { "publisher": "HuggingFaceTB", "document": "config.json vision_config", "doc_id": "hf:HuggingFaceTB/SmolVLM-Instruct", "url": "https://huggingface.co/HuggingFaceTB/SmolVLM-Instruct/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"hidden_size\": 1152, \"num_hidden_layers\": 27, \"image_size\": 384" } ], "mmproj": { "format": "GGUF mmproj Q8_0 (vision)", "file": "mmproj-SmolVLM-Instruct-Q8_0.gguf", "size_gb": 0.593, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "SmolVLM-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/SmolVLM-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "mmproj-SmolVLM-Instruct-Q8_0.gguf size 592521344 bytes" } ] } }, "image_tokens": { "tokens": 81, "class": "D", "basis": "384² patch / 14 ≈ 27² tokens, pixel-shuffle ×3 → 81 tokens per 384×384 tile; one tile assumed" }, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M (text)", "file": "SmolVLM-Instruct-Q4_K_M.gguf", "size_gb": 1.112, "sources": [ { "publisher": "ggml-org (llama.cpp maintainers)", "document": "SmolVLM-Instruct-GGUF repo file listing", "doc_id": "hf:ggml-org/SmolVLM-Instruct-GGUF", "url": "https://huggingface.co/api/models/ggml-org/SmolVLM-Instruct-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "SmolVLM-Instruct-Q4_K_M.gguf size 1112242368 bytes" } ] } ], "jetson_measurements": [ { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark, listed as 'SmolVLM 2B')", "quant": null, "memory_gb": null, "tokens_per_s": 12.9, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "SmolVLM 2B: Jetson Orin Nano 8.1, Jetson Orin Nano Super 12.9 (perf gain 1.59x)" } ] } ], "notes": "Model card evaluation table lists 'Min GPU RAM required (GB)' for SmolVLM (this 2.2B model) as 5.02 (generic GPU, not Jetson-specific)." }, "llava-1.5-7b-hf": { "label": "LLaVA 1.5 7B", "family": "llava", "type": "vlm", "hf_repo": "llava-hf/llava-1.5-7b-hf", "params_b": 7.063, "text_params_b": 6.761, "params_sources": [ { "publisher": "Hugging Face", "document": "model.safetensors.index.json", "doc_id": "hf:llava-hf/llava-1.5-7b-hf", "url": "https://huggingface.co/llava-hf/llava-1.5-7b-hf/raw/main/model.safetensors.index.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "\"total_size\": 14126854144" } ], "architecture": { "layers": 32, "hidden": 4096, "heads": 32, "kv_heads": 32, "head_dim": 128, "max_ctx": 4096, "dtype": "float16", "vocab": 32064, "head_dim_derived": false, "sources": [ { "publisher": "Hugging Face / lmsys", "document": "llava-1.5-7b-hf config.json text_config (backbone = lmsys/vicuna-7b-v1.5); explicit layer/head counts confirmed via lmsys/vicuna-7b-v1.5 config.json since llava-hf's text_config omits fields matching LlamaConfig defaults", "doc_id": "hf:llava-hf/llava-1.5-7b-hf", "url": "https://huggingface.co/llava-hf/llava-1.5-7b-hf/raw/main/config.json ; https://huggingface.co/lmsys/vicuna-7b-v1.5/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "text_config._name_or_path: \"lmsys/vicuna-7b-v1.5\", max_position_embeddings: 4096 (llava-1.5-7b-hf); \"num_attention_heads\": 32, \"num_hidden_layers\": 32, \"num_key_value_heads\": 32 (vicuna-7b-v1.5)" } ] }, "vision": { "layers": 24, "hidden": 1024, "patch_size": 14, "image_size": 336, "model_type": "clip_vision_model", "sources": [ { "publisher": "Hugging Face", "document": "config.json vision_config", "doc_id": "hf:llava-hf/llava-1.5-7b-hf", "url": "https://huggingface.co/llava-hf/llava-1.5-7b-hf/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"model_type\": \"clip_vision_model\", \"hidden_size\": 1024, \"image_size\": 336" } ], "mmproj": { "format": "GGUF mmproj f16 (vision)", "file": "llava-v1.5-7b-mmproj-model-f16.gguf", "size_gb": 0.624, "sources": [ { "publisher": "second-state (community GGUF)", "document": "Llava-v1.5-7B-GGUF repo file listing", "doc_id": "hf:second-state/Llava-v1.5-7B-GGUF", "url": "https://huggingface.co/api/models/second-state/Llava-v1.5-7B-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "llava-v1.5-7b-mmproj-model-f16.gguf size 624434368 bytes" } ] } }, "image_tokens": { "tokens": 576, "class": "D", "basis": "(336 / 14)² = 576 patches, no pooling (vision_config image_size 336, patch_size 14)" }, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M (text)", "file": "llava-v1.5-7b-Q4_K_M.gguf", "size_gb": 4.081, "sources": [ { "publisher": "second-state (community GGUF)", "document": "Llava-v1.5-7B-GGUF repo file listing", "doc_id": "hf:second-state/Llava-v1.5-7B-GGUF", "url": "https://huggingface.co/api/models/second-state/Llava-v1.5-7B-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "llava-v1.5-7b-Q4_K_M.gguf size 4081004320 bytes" } ] } ], "jetson_measurements": [ { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark, listed as 'LLAVA 1.6 7B', not 1.5)", "quant": null, "memory_gb": null, "tokens_per_s": 0.57, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "LLAVA 1.6 7B: Jetson Orin Nano 0.412, Jetson Orin Nano Super 0.57 (perf gain 1.38x)" } ] } ], "notes": "The archived Jetson benchmark table lists 'LLAVA 1.6 7B' (not 1.5) tokens/sec; recorded here as the closest architecturally-comparable published Jetson figure for a 7B LLaVA, flagged as a version mismatch. The tutorial page confirms a plain 'Llava-7b' (1.5 architecture) fits on Orin Nano 8GB with 4-bit quantization but does not give an exact GB figure." }, "vila-1.5-3b": { "label": "VILA 1.5 3B", "family": "vila", "type": "vlm", "hf_repo": "Efficient-Large-Model/VILA1.5-3b", "params_b": 3.148, "text_params_b": 2.718, "params_sources": [ { "publisher": "Hugging Face", "document": "repo blob listing (llm/, vision_tower/, mm_projector/ safetensors sizes)", "doc_id": "hf:Efficient-Large-Model/VILA1.5-3b", "url": "https://huggingface.co/api/models/Efficient-Large-Model/VILA1.5-3b?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "llm/model-00001-of-00002.safetensors 4974521464, llm/model-00002-of-00002.safetensors 428632856, vision_tower/model.safetensors 856506120, mm_projector/model.safetensors 36729360" } ], "architecture": { "layers": 32, "hidden": 2560, "heads": 20, "kv_heads": 20, "head_dim": 128, "max_ctx": 4096, "dtype": "bfloat16", "vocab": 32000, "head_dim_derived": false, "sources": [ { "publisher": "Efficient-Large-Model (VILA)", "document": "llm/config.json", "doc_id": "hf:Efficient-Large-Model/VILA1.5-3b", "url": "https://huggingface.co/Efficient-Large-Model/VILA1.5-3b/raw/main/llm/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"hidden_size\": 2560, \"num_attention_heads\": 20, \"num_key_value_heads\": 20, \"num_hidden_layers\": 32" } ] }, "vision": { "layers": 27, "hidden": 1152, "patch_size": 14, "image_size": 384, "model_type": "siglip_vision_model", "sources": [ { "publisher": "Efficient-Large-Model (VILA)", "document": "vision_tower/config.json", "doc_id": "hf:Efficient-Large-Model/VILA1.5-3b", "url": "https://huggingface.co/Efficient-Large-Model/VILA1.5-3b/raw/main/vision_tower/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"model_type\": \"siglip_vision_model\", \"hidden_size\": 1152, \"num_hidden_layers\": 27" } ] }, "image_tokens": { "tokens": 196, "class": "E", "basis": "SigLIP 384 / 14 ≈ 27² patches downsampled ~4× by the VILA projector; 196 assumed" }, "quantized_artefacts": [], "jetson_measurements": [ { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark)", "quant": null, "memory_gb": null, "tokens_per_s": 1.06, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "VILA 1.5 3B: Jetson Orin Nano 0.7, Jetson Orin Nano Super 1.06 (perf gain 1.51x)" } ] } ], "notes": "No GGUF quantization found for VILA (not supported by mainline llama.cpp as of retrieval date) -- see gaps. LLM backbone is a 2.7B-param Llama-architecture model (not a named Llama release); '3b' in the model name refers to the combined llm+vision+projector size (~3.15B computed here)." }, "vila-1.5-8b": { "label": "VILA 1.5 8B (Llama-3)", "family": "vila", "type": "vlm", "hf_repo": "Efficient-Large-Model/Llama-3-VILA1.5-8B", "params_b": 8.494, "text_params_b": 8.064, "params_sources": [ { "publisher": "Hugging Face", "document": "repo blob listing (llm/, vision_tower/, mm_projector/ safetensors sizes)", "doc_id": "hf:Efficient-Large-Model/Llama-3-VILA1.5-8B", "url": "https://huggingface.co/api/models/Efficient-Large-Model/Llama-3-VILA1.5-8B?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "llm/model-0000{1..4}-of-00004.safetensors sizes 4976706864 + 4999802720 + 4915916176 + 1168147000, vision_tower/model.safetensors 856506120, mm_projector/model.safetensors 71338512" } ], "architecture": { "layers": 32, "hidden": 4096, "heads": 32, "kv_heads": 8, "head_dim": 128, "max_ctx": 8192, "dtype": "bfloat16", "vocab": 128257, "head_dim_derived": false, "sources": [ { "publisher": "Efficient-Large-Model (VILA)", "document": "llm/config.json", "doc_id": "hf:Efficient-Large-Model/Llama-3-VILA1.5-8B", "url": "https://huggingface.co/Efficient-Large-Model/Llama-3-VILA1.5-8B/raw/main/llm/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"hidden_size\": 4096, \"num_attention_heads\": 32, \"num_key_value_heads\": 8, \"num_hidden_layers\": 32" } ] }, "vision": { "layers": 27, "hidden": 1152, "patch_size": 14, "image_size": 384, "model_type": "siglip_vision_model", "sources": [ { "publisher": "Efficient-Large-Model (VILA)", "document": "vision_tower/config.json", "doc_id": "hf:Efficient-Large-Model/Llama-3-VILA1.5-8B", "url": "https://huggingface.co/Efficient-Large-Model/Llama-3-VILA1.5-8B/raw/main/vision_tower/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "vision_config", "quote": "\"model_type\": \"siglip_vision_model\", \"hidden_size\": 1152, \"num_hidden_layers\": 27" } ] }, "image_tokens": { "tokens": 196, "class": "E", "basis": "SigLIP 384 / 14 ≈ 27² patches downsampled ~4× by the VILA projector; 196 assumed" }, "quantized_artefacts": [], "jetson_measurements": [ { "module": "jetson_orin_nano_super", "runtime": "unspecified (archived benchmark)", "quant": null, "memory_gb": null, "tokens_per_s": 0.83, "sources": [ { "publisher": "NVIDIA Jetson AI Lab (archive)", "document": "benchmarks.html", "doc_id": "jetson-ai-lab:benchmarks.html", "url": "https://www.jetson-ai-lab.com/archive/benchmarks.html", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "C", "quote": "VILA 1.5 8B: Jetson Orin Nano 0.574, Jetson Orin Nano Super 0.83 (perf gain 1.45x)" } ] } ], "notes": "No GGUF quantization found for VILA -- see gaps. LLM backbone is Llama-3-8B architecture (matches llama-3.1-8b-instruct layer/head config except max_position_embeddings=8192 here vs 131072 for the released Llama 3.1 chat model, and rope_theta 500000)." }, "whisper-small": { "label": "Whisper small", "family": "whisper", "type": "asr", "hf_repo": "openai/whisper-small", "params_b": 0.244, "text_params_b": 0.244, "params_sources": [ { "publisher": "OpenAI", "document": "openai/whisper GitHub README.md", "doc_id": "github:openai/whisper/README.md", "url": "https://raw.githubusercontent.com/openai/whisper/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "| small | 244 M | `small.en` | `small` | ~2 GB | ~4x |" } ], "architecture": { "layers": 12, "hidden": 768, "heads": 12, "kv_heads": 12, "head_dim": 64, "max_ctx": 448, "dtype": "float32", "vocab": 51865, "head_dim_derived": false, "sources": [ { "publisher": "OpenAI", "document": "config.json", "doc_id": "hf:openai/whisper-small", "url": "https://huggingface.co/openai/whisper-small/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"d_model\": 768, \"encoder_layers\": 12, \"decoder_layers\": 12, \"encoder_attention_heads\": 12, \"decoder_attention_heads\": 12" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [], "jetson_measurements": [], "notes": "Encoder-decoder architecture: num_hidden_layers/hidden_size/heads above are per-stack (encoder_layers==decoder_layers==12, encoder_attention_heads==decoder_attention_heads==12 for this size). max_position_embeddings=448 is max_target_positions (decoder text token limit); encoder audio context is max_source_positions=1500 (30s of 20ms mel frames). torch_dtype in config.json is float32 (not bf16) -- params computed from OpenAI's published table, not from safetensors size, per instructions (parameters only for Whisper). Reference VRAM requirement '~2 GB' from OpenAI README is a generic estimate, not Jetson-measured." }, "whisper-medium": { "label": "Whisper medium", "family": "whisper", "type": "asr", "hf_repo": "openai/whisper-medium", "params_b": 0.769, "text_params_b": 0.769, "params_sources": [ { "publisher": "OpenAI", "document": "openai/whisper GitHub README.md", "doc_id": "github:openai/whisper/README.md", "url": "https://raw.githubusercontent.com/openai/whisper/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "| medium | 769 M | `medium.en` | `medium` | ~5 GB | ~2x |" } ], "architecture": { "layers": 24, "hidden": 1024, "heads": 16, "kv_heads": 16, "head_dim": 64, "max_ctx": 448, "dtype": "float32", "vocab": 51865, "head_dim_derived": false, "sources": [ { "publisher": "OpenAI", "document": "config.json", "doc_id": "hf:openai/whisper-medium", "url": "https://huggingface.co/openai/whisper-medium/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"d_model\": 1024, \"encoder_layers\": 24, \"decoder_layers\": 24, \"encoder_attention_heads\": 16, \"decoder_attention_heads\": 16" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [], "jetson_measurements": [], "notes": "Encoder-decoder architecture; fields per-stack (encoder==decoder here). Reference VRAM requirement '~5 GB' from OpenAI README is a generic estimate, not Jetson-measured." }, "whisper-large-v3": { "label": "Whisper large-v3", "family": "whisper", "type": "asr", "hf_repo": "openai/whisper-large-v3", "params_b": 1.55, "text_params_b": 1.55, "params_sources": [ { "publisher": "OpenAI", "document": "openai/whisper GitHub README.md", "doc_id": "github:openai/whisper/README.md", "url": "https://raw.githubusercontent.com/openai/whisper/main/README.md", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "| large | 1550 M | N/A | `large` | ~10 GB | 1x |" } ], "architecture": { "layers": 32, "hidden": 1280, "heads": 20, "kv_heads": 20, "head_dim": 64, "max_ctx": 448, "dtype": "float16", "vocab": 51866, "head_dim_derived": false, "sources": [ { "publisher": "OpenAI", "document": "config.json", "doc_id": "hf:openai/whisper-large-v3", "url": "https://huggingface.co/openai/whisper-large-v3/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"d_model\": 1280, \"encoder_layers\": 32, \"decoder_layers\": 32, \"encoder_attention_heads\": 20, \"decoder_attention_heads\": 20, \"num_mel_bins\": 128" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [], "jetson_measurements": [], "notes": "large-v3 uses 128 mel bins (vs 80 for small/medium/large-v2). Reference VRAM requirement '~10 GB' from OpenAI README is a generic estimate, not Jetson-measured. HF repo also hosts fp16 model.safetensors (1543130976 bytes) confirming ~1.54-1.55B param scale." }, "mistral-7b-instruct-v0.3": { "label": "Mistral 7B Instruct v0.3", "family": "mistral", "type": "llm", "hf_repo": "mistralai/Mistral-7B-Instruct-v0.3", "params_b": 7.248, "text_params_b": 7.248, "params_sources": [ { "publisher": "Hugging Face", "document": "model.safetensors.index.json", "doc_id": "hf:mistralai/Mistral-7B-Instruct-v0.3", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/raw/main/model.safetensors.index.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "\"total_size\": 14496047104" } ], "architecture": { "layers": 32, "hidden": 4096, "heads": 32, "kv_heads": 8, "head_dim": 128, "max_ctx": 32768, "dtype": "bfloat16", "vocab": 32768, "head_dim_derived": false, "sources": [ { "publisher": "Mistral AI", "document": "config.json", "doc_id": "hf:mistralai/Mistral-7B-Instruct-v0.3", "url": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/raw/main/config.json", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "section": "config.json", "quote": "\"num_attention_heads\": 32, \"num_key_value_heads\": 8, \"vocab_size\": 32768" } ] }, "vision": null, "image_tokens": null, "quantized_artefacts": [ { "quant": "q4", "format": "GGUF Q4_K_M", "file": "Mistral-7B-Instruct-v0.3-Q4_K_M.gguf", "size_gb": 4.373, "sources": [ { "publisher": "bartowski (community GGUF)", "document": "Mistral-7B-Instruct-v0.3-GGUF repo file listing", "doc_id": "hf:bartowski/Mistral-7B-Instruct-v0.3-GGUF", "url": "https://huggingface.co/api/models/bartowski/Mistral-7B-Instruct-v0.3-GGUF?blobs=true", "retrieved": "2026-09-07", "verified": "2026-09-07", "class": "A", "quote": "Mistral-7B-Instruct-v0.3-Q4_K_M.gguf size 4372812000 bytes" } ] } ], "jetson_measurements": [], "notes": "head_dim derived as hidden_size/num_attention_heads=4096/32=128 (no explicit head_dim field in config.json). params_b (7.25B) is higher than the nominal '7B' name due to the extended 32768-token vocabulary vs Mistral-7B-v0.1's 32000." } } }