VILA 1.5 3B: memory fit on every Jetson module.
Method v1.0 · dataset 2026-09-07 · verified 2026-09-07 · last updated September 2026
3.15 B parameters (text 2.72 B + vision tower) · 32 layers · 4,096-token max context. Hugging Face: Efficient-Large-Model/VILA1.5-3b.
Model facts
| Fact | Value | Note |
|---|---|---|
| Parameters | 3.148 B | text 2.718 B + vision 0.43 B (estimated) |
| Layers | 32 | |
| Hidden size | 2560 | |
| Attention heads | 20 | |
| KV heads | 20 | |
| Head dimension | 128 | |
| Max context | 4,096 tokens | |
| dtype | bfloat16 | |
| Vocabulary | 32,000 |
Efficient-Large-Model (VILA) · hf:Efficient-Large-Model/VILA1.5-3b · config.json · verified 2026-09-07
“"hidden_size": 2560, "num_attention_heads": 20, "num_key_value_heads": 20, "num_hidden_layers": 32”
Parameters source
- Hugging Face · hf:Efficient-Large-Model/VILA1.5-3b · verified 2026-09-07 · repo blob listing (llm/, vision_tower/, mm_projector/ safetensors sizes) · class A
“llm/model-00001-of-00002.safetensors 4974521464, llm/model-00002-of-00002.safetensors 428632856, vision_tower/model.safetensors 856506120, mm_projector/model.safetensors 36729360”
Vision tower
| Fact | Value |
|---|---|
| Layers | 27 |
| Hidden size | 1152 |
| Patch size | 14 |
| Image size | 384 |
| Model type | siglip_vision_model |
Vision config source
- Efficient-Large-Model (VILA) · hf:Efficient-Large-Model/VILA1.5-3b · vision_config · verified 2026-09-07 · vision_tower/config.json · class A
“"model_type": "siglip_vision_model", "hidden_size": 1152, "num_hidden_layers": 27”
Image tokens: 196 tokens per image (SigLIP 384 / 14 ≈ 27² patches downsampled ~4× by the VILA projector; 196 assumed; class E).
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).
No published quantised artefact for this model; weights are estimated as parameters × bytes-per-parameter for every quantisation.
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) | — | 1.06 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
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).
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.