LLaVA 1.5 7B: memory fit on every Jetson module.
Method v1.0 · dataset 2026-09-07 · verified 2026-09-07 · last updated September 2026
7.06 B parameters (text 6.76 B + vision tower) · 32 layers · 4,096-token max context. Hugging Face: llava-hf/llava-1.5-7b-hf.
Model facts
| Fact | Value | Note |
|---|---|---|
| Parameters | 7.063 B | text 6.761 B + vision 0.302 B (estimated) |
| Layers | 32 | |
| Hidden size | 4096 | |
| Attention heads | 32 | |
| KV heads | 32 | |
| Head dimension | 128 | |
| Max context | 4,096 tokens | |
| dtype | float16 | |
| Vocabulary | 32,064 |
Hugging Face / lmsys · hf:llava-hf/llava-1.5-7b-hf · config.json · verified 2026-09-07
“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)”
Parameters source
- Hugging Face · hf:llava-hf/llava-1.5-7b-hf · verified 2026-09-07 · model.safetensors.index.json · class A
“"total_size": 14126854144”
Vision tower
| Fact | Value |
|---|---|
| Layers | 24 |
| Hidden size | 1024 |
| Patch size | 14 |
| Image size | 336 |
| Model type | clip_vision_model |
Vision config source
- Hugging Face · hf:llava-hf/llava-1.5-7b-hf · vision_config · verified 2026-09-07 · config.json vision_config · class A
“"model_type": "clip_vision_model", "hidden_size": 1024, "image_size": 336”
mmproj (vision weights): GGUF mmproj f16 (vision), 0.624 GB.
mmproj source
- second-state (community GGUF) · hf:second-state/Llava-v1.5-7B-GGUF · verified 2026-09-07 · Llava-v1.5-7B-GGUF repo file listing · class A
“llava-v1.5-7b-mmproj-model-f16.gguf size 624434368 bytes”
Image tokens: 576 tokens per image ((336 / 14)² = 576 patches, no pooling (vision_config image_size 336, patch_size 14); class D).
Published quantised artefacts
A file the publisher or a community mirror actually ships, used as the class-A weight figure when the requested quant matches. Any quant without a row here falls back to parameters × bytes-per-parameter (class D for block quants).
| Quant | Format | File | Size | Source |
|---|---|---|---|---|
| GGUF Q4_K_M / INT4 (AWQ, MLC q4f16) | GGUF Q4_K_M (text) | llava-v1.5-7b-Q4_K_M.gguf | 4.081 GB | second-state (community GGUF) · hf:second-state/Llava-v1.5-7B-GGUF · verified 2026-09-07 |
Jetson tokens/s measurements
Published or archived throughput numbers, not modelled. Class C (external measured benchmark).
| Module | Runtime | Quant | Measured | Source |
|---|---|---|---|---|
| Jetson Orin Nano Super | unspecified (archived benchmark, listed as 'LLAVA 1.6 7B', not 1.5) | — | 0.57 tok/s | NVIDIA Jetson AI Lab (archive) · jetson-ai-lab:benchmarks.html · verified 2026-09-07 |
Every module × every quantisation
Verdict and total memory at 4096-token context, 1 sequence, llama.cpp, headless. Q4 cells link to the static breakdown page; every other cell links to the live engine at that quantisation.
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.
Method and data: Model Memory Fit methodology. Full registry: model-memory.json.
Change the context, concurrency or runtime.
The live engine covers any context length, sequence count, KV precision and runtime, with a permanent link.