Gemma 3 4B IT: memory fit on every Jetson module.
Method v1.0 · dataset 2026-09-07 · verified 2026-09-07 · last updated September 2026
4.3 B parameters (text 3.87 B + vision tower) · 34 layers · 131,072-token max context. Hugging Face: google/gemma-3-4b-it.
Model facts
| Fact | Value | Note |
|---|---|---|
| Parameters | 4.3 B | text 3.87 B + vision 0.43 B (estimated) |
| Layers | 34 | |
| Hidden size | 2560 | |
| Attention heads | 8 | |
| KV heads | 4 | |
| Head dimension | 256 | |
| Max context | 131,072 tokens | |
| dtype | bfloat16 | |
| Vocabulary | 262,208 |
Hugging Face (mirror: unsloth/gemma-3-4b-it) · hf:unsloth/gemma-3-4b-it · config.json · verified 2026-09-07
“"num_attention_heads": 8, "num_key_value_heads": 4, "head_dim": 256”
Parameters source
- Hugging Face (mirror: unsloth/gemma-3-4b-it, identical weights to google/gemma-3-4b-it) · hf:unsloth/gemma-3-4b-it · verified 2026-09-07 · model.safetensors.index.json · class A
“"total_size": 8600158944”
Vision tower
| Fact | Value |
|---|---|
| Layers | 27 |
| Hidden size | 1152 |
| Patch size | 14 |
| Image size | 896 |
| Model type | siglip_vision_model |
Vision config source
- Hugging Face (mirror: unsloth/gemma-3-4b-it) · hf:unsloth/gemma-3-4b-it · vision_config · verified 2026-09-07 · config.json vision_config · class A
“"model_type": "siglip_vision_model", "hidden_size": 1152, "image_size": 896”
Image tokens: 256 tokens per image (SigLIP 896 / 14 = 64² patches pooled to 256 soft tokens per image (Gemma 3 technical report §2); 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 | gemma-3-4b-it-Q4_K_M.gguf | 2.49 GB | unsloth (community GGUF) · hf:unsloth/gemma-3-4b-it-GGUF · verified 2026-09-07 |
Jetson tokens/s measurements
Published or archived throughput numbers, not modelled. Class C (external measured benchmark).
No Jetson-measured tokens/s reported for this model yet.
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
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).
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.