Llama 3.2 3B Instruct
Meta
Small enough for a laptop CPU or a 6 GB card, and still coherent. A sensible target for edge deployment.
Excellent fit
- Weights
- 6.0 GB
- KV cache
- 880 MB
- Overhead
- 510 MB
- 7.3 GB of 31 GB usable VRAM.
- Chosen as the best quality that still fits a 32,768-token context (10.0 GB at that length).
- Runs almost anywhere
- Very fast
- Cheap to fine-tune
- Limited reasoning depth
- Hallucinates more than 7B+ models
llama3.2:3bRunning Llama 3.2 3B Instruct on GeForce RTX 5090
| Quantization | Quality | Weights | KV cache | Total | ~tok/s | Fit |
|---|---|---|---|---|---|---|
| F16Pick | lossless | 6.0 GB | 880 MB | 7.3 GB | 216 | Excellent fit |
| Q8_0 | near-lossless | 3.2 GB | 880 MB | 4.6 GB | 407 | Excellent fit |
| Q6_K | near-lossless | 2.4 GB | 880 MB | 3.8 GB | 528 | Excellent fit |
| Q5_K_M | high | 2.1 GB | 880 MB | 3.5 GB | 611 | Excellent fit |
| Q4_K_M | balanced | 1.8 GB | 880 MB | 3.2 GB | 717 | Excellent fit |
| Q3_K_M | degraded | 1.5 GB | 880 MB | 2.9 GB | 886 | 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
- 1.7 GB
- Optimizer
- 362 MB
- Activations
- 570 MB
- Peak
- 3.9 GB
- 3.9 GB peak against 31 GB usable — room to raise batch size or sequence length.
- Base weights
- 6.0 GB
- Optimizer
- 362 MB
- Activations
- 570 MB
- Peak
- 8.2 GB
- 8.2 GB peak against 31 GB usable — room to raise batch size or sequence length.
Where this model runs
VRAM 32 GB · F16 · 7.3 GB
VRAM 24 GB · F16 · 7.3 GB
VRAM 16 GB · F16 · 7.3 GB
VRAM 16 GB · F16 · 7.3 GB
VRAM 24 GB · F16 · 7.3 GB
VRAM 12 GB · Q8_0 · 4.6 GB
VRAM 12 GB · Q8_0 · 4.6 GB
VRAM 16 GB · F16 · 7.3 GB
VRAM 48 GB · F16 · 7.3 GB
VRAM 80 GB · F16 · 7.3 GB
Unified 128 GB · F16 · 7.3 GB
Unified 48 GB · F16 · 7.3 GB
Unified 24 GB · F16 · 7.3 GB
Unified 192 GB · F16 · 7.3 GB
Unified 16 GB · Q8_0 · 4.6 GB
VRAM 24 GB · F16 · 7.3 GB
VRAM 16 GB · F16 · 7.3 GB
VRAM 0 MB · Q4_K_M · 3.2 GB
VRAM 0 MB · Q4_K_M · 3.2 GB
63.4%
Llama 3.2 model card
Reported by the model's author. ModelLM has not run these benchmarks and does not treat them as verified.
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.