unsloth
Ministral 3 8B Reasoning 2512
Ministral 3 8B Reasoning 2512 is a 8.49B dense mistral3 model from unsloth. Configured context length is 262,144 tokens. License: apache-2.0. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters8.5B
Context256K
Attentiongqa
Licenseapache-2.0
34 layers · 8 KV heads · head dim 128 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 15.8 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 10.3 GB | Hugging Face file bytes |
| Q8_0 | 8.41 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 7.19 GB | Hugging Face file bytes |
| Q6_K | 6.49 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 5.66 GB | Hugging Face file bytes |
| Q5_K_M | 5.64 GB | Hugging Face file bytes |
| Q5_K_S | 5.51 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 4.92 GB | Hugging Face file bytes |
| Q4_K_M | 4.84 GB | Hugging Face file bytes |
| Q4_K_S | 4.61 GB | Hugging Face file bytes |
| IQ4_NL | 4.60 GB | Hugging Face file bytes |
| IQ4_XS | 4.39 GB | Hugging Face file bytes |
| Q4_1 | 5.05 GB | Hugging Face file bytes |
| Q4_0 | 4.60 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 4.12 GB | Hugging Face file bytes |
| Q3_K_M | 3.95 GB | Hugging Face file bytes |
| Q3_K_S | 3.60 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 3.26 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 3.32 GB | Hugging Face file bytes |
| Q2_K | 6.36 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 2.45 GB | Hugging Face file bytes |
| UD-IQ1_M | 2.24 GB | Hugging Face file bytes |
| UD-IQ1_S | 2.11 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Tight fit | UD-Q5_K_XL | 24–33 |
| GeForce RTX 3060 12GB | Tight fit | Q8_0 | 19–27 |
| GeForce RTX 4060 Ti 16GB | Runs | UD-Q8_K_XL | 13–18 |
| GeForce RTX 4090 | Runs | BF16 | 31–42 |
| GeForce RTX 5090 | Runs great | BF16 | 54–74 |
| Apple M4 Pro | Tight fit | BF16 | 8.3–11 |
| Apple M5 | Tight fit | Q8_0 | 8.3–11 |
| CPU only · DDR5-5600 | Runs on CPU | BF16 | 2.7–3.7 |
Smallest devices that fit it fully
- GeForce RTX 3050 6GB — Tight fit, UD-Q3_K_XL, 18–25 tok/s
- GeForce RTX 2060 6GB — Tight fit, UD-Q3_K_XL, 37–50 tok/s
- GeForce GTX 1660 Super — Tight fit, UD-Q3_K_XL, 37–50 tok/s
- GeForce GTX 1660 Ti — Tight fit, UD-Q3_K_XL, 32–43 tok/s
- Intel Arc A380 — Tight fit, UD-Q3_K_XL, 20–28 tok/s
- GeForce RTX 5060 Ti 8GB — Tight fit, UD-Q5_K_XL, 37–51 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (UD-Q8_K_XL).
llama.cpp
llama-server -hf unsloth/Ministral-3-8B-Reasoning-2512-GGUF:UD-Q8_K_XL -c 8192 -ngl 34
No matching name in the Ollama library snapshot.
LM Studio
lms get unsloth/Ministral-3-8B-Reasoning-2512-GGUF # Then load quant UD-Q8_K_XL