Mistral Nemo 12B Instruct
Mistral AI / NVIDIA
A 12B with a 128k context and the Tekken tokenizer, which compresses non-English text far better than Llama’s.
Good fit
- Weights
- 8.1 GB
- KV cache
- 1.6 GB
- Overhead
- 670 MB
- 10 GB of 15 GB usable VRAM.
- Chosen as the best quality that still fits a 16,384-token context (12 GB at that length).
- 128k context
- Efficient multilingual tokenizer
- Apache 2.0
- Middling reasoning for its size
- Long context needs a large KV cache
mistral-nemo:12bRunning Mistral Nemo 12B Instruct on GeForce RTX 4060 Ti 16 GB
| Quantization | Quality | Weights | KV cache | Total | ~tok/s | Fit |
|---|---|---|---|---|---|---|
| F16 | lossless | 23 GB | 1.6 GB | 25 GB | 9 | Runs with CPU offload |
| Q8_0 | near-lossless | 12 GB | 1.6 GB | 14 GB | 17 | Runs with CPU offload |
| Q6_K | near-lossless | 9.3 GB | 1.6 GB | 12 GB | 22 | Good fit |
| Q5_K_MPick | high | 8.1 GB | 1.6 GB | 10 GB | 26 | Good fit |
| Q4_K_M | balanced | 6.9 GB | 1.6 GB | 9.1 GB | 30 | Good fit |
| Q3_K_M | degraded | 5.5 GB | 1.6 GB | 7.8 GB | 37 | 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
- 6.4 GB
- Optimizer
- 874 MB
- Activations
- 1.2 GB
- Peak
- 10 GB
- 10 GB peak against 15 GB usable — room to raise batch size or sequence length.
- Base weights
- 23 GB
- Optimizer
- 874 MB
- Activations
- 1.2 GB
- Peak
- 26 GB
- Needs 26 GB — switch to QLoRA to cut the weight footprint.
Where this model runs
VRAM 32 GB · Q8_0 · 14 GB
VRAM 24 GB · Q6_K · 12 GB
VRAM 16 GB · Q5_K_M · 10 GB
VRAM 16 GB · Q5_K_M · 10 GB
VRAM 24 GB · Q6_K · 12 GB
VRAM 12 GB · Q4_K_M · 9.1 GB
VRAM 12 GB · Q4_K_M · 9.1 GB
VRAM 16 GB · Q5_K_M · 10 GB
VRAM 48 GB · F16 · 25 GB
VRAM 80 GB · F16 · 25 GB
Unified 128 GB · F16 · 25 GB
Unified 48 GB · Q8_0 · 14 GB
Unified 24 GB · Q4_K_M · 9.1 GB
Unified 192 GB · F16 · 25 GB
Unified 16 GB · Q4_K_M · 9.1 GB
VRAM 24 GB · Q6_K · 12 GB
VRAM 16 GB · Q5_K_M · 10 GB
VRAM 0 MB
VRAM 0 MB
68%
Mistral Nemo announcement
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 14B Instruct
14.8B · Apache 2.0
The sweet spot for 24 GB cards. Meaningfully stronger reasoning than 7B while still fine-tunable locally with QLoRA.
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.
Catalogue figures come from each model’s published card. ModelLM has not independently measured them.