Gemma 3 4B Instruct
Vision plus 128k context on a laptop. The most capable genuinely small multimodal option.
Excellent fit
- Weights
- 8.0 GB
- KV cache
- 1.3 GB
- Overhead
- 530 MB
- 9.9 GB of 31 GB usable VRAM.
- Chosen as the best quality that still fits a 32,768-token context (14 GB at that length).
- Multimodal at 4B
- Runs on 8 GB cards and M-series laptops
- Long context
- Limited reasoning
- Vocabulary dominates the parameter count at this size
gemma3:4bRunning Gemma 3 4B Instruct on GeForce RTX 5090
| Quantization | Quality | Weights | KV cache | Total | ~tok/s | Fit |
|---|---|---|---|---|---|---|
| F16Pick | lossless | 8.0 GB | 1.3 GB | 9.9 GB | 161 | Excellent fit |
| Q8_0 | near-lossless | 4.3 GB | 1.3 GB | 6.1 GB | 303 | Excellent fit |
| Q6_K | near-lossless | 3.3 GB | 1.3 GB | 5.1 GB | 393 | Excellent fit |
| Q5_K_M | high | 2.8 GB | 1.3 GB | 4.7 GB | 455 | Excellent fit |
| Q4_K_M | balanced | 2.4 GB | 1.3 GB | 4.3 GB | 534 | Excellent fit |
| Q3_K_M | degraded | 2.0 GB | 1.3 GB | 3.8 GB | 659 | Excellent fit |
KV cache is sized at 8,192 tokens. Longer contexts cost proportionally more — the recommendation above reserves room for a working context.
You can fine-tune this here
- Base weights
- 2.3 GB
- Optimizer
- 457 MB
- Activations
- 610 MB
- Peak
- 4.7 GB
- 4.7 GB peak against 31 GB usable — room to raise batch size or sequence length.
- Base weights
- 8.0 GB
- Optimizer
- 457 MB
- Activations
- 610 MB
- Peak
- 10 GB
- 10 GB peak against 31 GB usable — room to raise batch size or sequence length.
Where this model runs
VRAM 32 GB · F16 · 9.9 GB
VRAM 24 GB · F16 · 9.9 GB
VRAM 16 GB · Q8_0 · 6.1 GB
VRAM 16 GB · Q8_0 · 6.1 GB
VRAM 24 GB · F16 · 9.9 GB
VRAM 12 GB · Q5_K_M · 4.7 GB
VRAM 12 GB · Q5_K_M · 4.7 GB
VRAM 16 GB · Q8_0 · 6.1 GB
VRAM 48 GB · F16 · 9.9 GB
VRAM 80 GB · F16 · 9.9 GB
Unified 128 GB · F16 · 9.9 GB
Unified 48 GB · F16 · 9.9 GB
Unified 24 GB · F16 · 9.9 GB
Unified 192 GB · F16 · 9.9 GB
Unified 16 GB · Q6_K · 5.1 GB
VRAM 24 GB · F16 · 9.9 GB
VRAM 16 GB · Q8_0 · 6.1 GB
VRAM 0 MB
VRAM 0 MB
Qwen2.5 7B Instruct
7.6B · Apache 2.0
The default starting point for local work on 8–12 GB cards. Strong instruction following and reliable tool-call formatting for its size.
Qwen2.5 Coder 7B Instruct
7.6B · Apache 2.0
The practical local copilot. Supports fill-in-the-middle, so it works as an inline completion model rather than only a chat assistant.
Qwen3 8B
8.2B · Apache 2.0
Switchable thinking mode: the same weights answer directly or reason step by step depending on the prompt. Long context for its size.
Llama 3.1 8B Instruct
8B · Llama 3.1 Community License
The most widely supported open model there is. If a tool, adapter or tutorial exists, it was written for this one first.
Catalogue figures come from each model’s published card. ModelLM has not independently measured them.