unsloth
Devstral 2 123B Instruct 2512
Devstral 2 123B Instruct 2512 is a 125.0B dense llama model from unsloth. Configured context length is 262,144 tokens. License: other. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters125B
Context256K
Attentiongqa
Licenseother
88 layers · 8 KV heads · head dim 128 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 232.9 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 136.0 GB | Hugging Face file bytes |
| Q8_0 | 123.7 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 102.3 GB | Hugging Face file bytes |
| Q6_K | 95.5 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 82.2 GB | Hugging Face file bytes |
| Q5_K_M | 82.3 GB | Hugging Face file bytes |
| Q5_K_S | 80.3 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 70.3 GB | Hugging Face file bytes |
| Q4_K_M | 69.8 GB | Hugging Face file bytes |
| Q4_K_S | 66.4 GB | Hugging Face file bytes |
| IQ4_NL | 66.0 GB | Hugging Face file bytes |
| IQ4_XS | 62.5 GB | Hugging Face file bytes |
| Q4_1 | 73.1 GB | Hugging Face file bytes |
| Q4_0 | 66.1 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 57.7 GB | Hugging Face file bytes |
| Q3_K_M | 56.5 GB | Hugging Face file bytes |
| Q3_K_S | 50.6 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 45.6 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 44.6 GB | Hugging Face file bytes |
| Q2_K | 87.1 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 32.0 GB | Hugging Face file bytes |
| UD-IQ1_M | 28.5 GB | Hugging Face file bytes |
| UD-IQ1_S | 26.5 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Partial offload | UD-IQ1_M | 1.8–2.5 |
| GeForce RTX 3060 12GB | Partial offload | UD-IQ2_XXS | 1.8–2.4 |
| GeForce RTX 4060 Ti 16GB | Partial offload | UD-IQ2_XXS | 1.9–2.6 |
| GeForce RTX 4090 | Partial offload | UD-IQ3_XXS | 1.7–2.3 |
| GeForce RTX 5090 | Tight fit | UD-IQ1_S | 33–45 |
| Apple M4 Pro | Won't fit | UD-IQ1_S | — |
| Apple M5 | Won't fit | UD-IQ1_S | — |
| CPU only · DDR5-5600 | Won't fit | UD-IQ1_S | — |
Smallest devices that fit it fully
- GeForce RTX 5090 — Tight fit, UD-IQ1_S, 33–45 tok/s
- RTX 5000 Ada — Tight fit, UD-IQ1_S, 11–14 tok/s
- Tesla V100 32GB — Tight fit, UD-IQ1_S, 17–23 tok/s
- NVIDIA A100 40GB — Tight fit, UD-IQ2_XXS, 23–31 tok/s
- NVIDIA L40S — Runs, UD-IQ2_XXS, 13–17 tok/s
- NVIDIA A40 — Runs, UD-IQ2_XXS, 10–14 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (UD-IQ2_XXS).
llama.cpp
llama-server -hf unsloth/Devstral-2-123B-Instruct-2512-GGUF:UD-IQ2_XXS -c 4096 -ngl 35
Ollama
ollama run devstral
LM Studio
lms get unsloth/Devstral-2-123B-Instruct-2512-GGUF # Then load quant UD-IQ2_XXS