SmolVLM 2.2B Instruct: memory fit on every Jetson module.
Method v1.0 · dataset 2026-09-07 · verified 2026-09-07 · last updated September 2026
2.25 B parameters (text 1.82 B + vision tower) · 24 layers · 16,384-token max context. Hugging Face: HuggingFaceTB/SmolVLM-Instruct.
Model facts
| Fact | Value | Note |
|---|---|---|
| Parameters | 2.246 B | text 1.816 B + vision 0.43 B (estimated) |
| Layers | 24 | |
| Hidden size | 2048 | |
| Attention heads | 32 | |
| KV heads | 32 | |
| Head dimension | 64 | |
| Max context | 16,384 tokens | |
| dtype | bfloat16 | |
| Vocabulary | 49,155 |
HuggingFaceTB · hf:HuggingFaceTB/SmolVLM-Instruct · config.json · verified 2026-09-07
“"num_attention_heads": 32, "num_key_value_heads": 32, "head_dim": 64”
Parameters source
- Hugging Face · hf:HuggingFaceTB/SmolVLM-Instruct · verified 2026-09-07 · model.safetensors (blob metadata via HF API) · class A
“model.safetensors size 4492630912 bytes”
Vision tower
| Fact | Value |
|---|---|
| Layers | 27 |
| Hidden size | 1152 |
| Patch size | 14 |
| Image size | 384 |
| Model type | idefics3 (SigLIP-SO400M vision encoder) |
Vision config source
- HuggingFaceTB · hf:HuggingFaceTB/SmolVLM-Instruct · vision_config · verified 2026-09-07 · config.json vision_config · class A
“"hidden_size": 1152, "num_hidden_layers": 27, "image_size": 384”
mmproj (vision weights): GGUF mmproj Q8_0 (vision), 0.593 GB.
mmproj source
- ggml-org (llama.cpp maintainers) · hf:ggml-org/SmolVLM-Instruct-GGUF · verified 2026-09-07 · SmolVLM-Instruct-GGUF repo file listing · class A
“mmproj-SmolVLM-Instruct-Q8_0.gguf size 592521344 bytes”
Image tokens: 81 tokens per image (384² patch / 14 ≈ 27² tokens, pixel-shuffle ×3 → 81 tokens per 384×384 tile; one tile assumed; 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) | SmolVLM-Instruct-Q4_K_M.gguf | 1.112 GB | ggml-org (llama.cpp maintainers) · hf:ggml-org/SmolVLM-Instruct-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 'SmolVLM 2B') | — | 12.9 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
Model card evaluation table lists 'Min GPU RAM required (GB)' for SmolVLM (this 2.2B model) as 5.02 (generic GPU, not Jetson-specific).
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.