readygpu

Ministral 3 8B Reasoning 2512

8.9B parameters · 256k context · Apache 2.0 · released 2025-10-31

chatcodingagenticvisionreasoning

In short

Ministral 3 8B Reasoning 2512 is an open model from Mistral AI with 8.9B parameters. At Q4_K_M, Ministral 3 8B Reasoning 2512 needs about 7 GB of memory at 8k context and 23 GB at 128k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the US that runs Ministral 3 8B Reasoning 2512 fully in memory at 128k context is the NVIDIA GeForce RTX 3090, $1,600 at eBay, at about 25 tok/s (estimated).

Updated . Speeds are estimates; how they are worked out.

What to run it on in the US

Fully in memory at 128k context, priced at retailers today.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA Tesla P40Q4_K_Mfits9 tok/s$34526.4
NVIDIA GeForce RTX 3090Q4_K_Mfits25 tok/s$1,60015.4
AMD Radeon AI PRO R9700Q8_0fits13 tok/s$1,7007.6
NVIDIA GeForce RTX 3090 TiQ4_K_Mfits27 tok/s$2,06012.9
NVIDIA GeForce RTX 4090Q4_K_Mfits27 tok/s$2,07312.8
Mac mini M4 Pro 48GBQ8_0fits7 tok/s$2,1493.0
Mac mini M4 Pro 64GBQ8_0fits7 tok/s$2,9002.2
Mac Studio M4 Max 64GBQ8_0fits13 tok/s$3,4003.8
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits5 tok/s$3,4491.5
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits5 tok/s$3,6501.4
Mac Studio M3 Ultra 96GBQ8_0fits20 tok/s$4,3994.4
NVIDIA GeForce RTX 5090Q8_0fits41 tok/s$4,9218.3
Mac Studio M4 Max 128GBQ8_0fits13 tok/s$4,9992.6
ASUS Ascent GX10 128GBQ8_0fits6 tok/s$6,1001.0
NVIDIA DGX Spark 128GBQ8_0fits6 tok/s$7,3990.8
NVIDIA RTX PRO 6000 BlackwellQ8_0fits41 tok/s$7,9975.1
Mac Studio M3 Ultra 256GBQ8_0fits20 tok/s$8,4992.3
Mac Studio M3 Ultra 512GBQ8_0fits20 tok/s$23,0000.8
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits5 tok/s––
Intel Arc B580Q4_K_Moffload4 tok/s$34610.1
NVIDIA GeForce RTX 3060 12GBQ4_K_Moffload4 tok/s$3609.7
Intel Arc A770 16GBQ4_K_Moffload5 tok/s$40011.5
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload3 tok/s$4697.2
AMD Radeon RX 7600 XTQ4_K_Moffload4 tok/s$4808.1
AMD Radeon RX 9060 XT 16GBQ4_K_Moffload4 tok/s$5207.9
AMD Radeon RX 7800 XTQ4_K_Moffload5 tok/s$5508.7
NVIDIA GeForce RTX 5060 Ti 16GBQ4_K_Moffload5 tok/s$6187.6
Intel Arc Pro B60Q4_K_Moffload10 tok/s$65014.6
NVIDIA GeForce RTX 4070 SuperQ4_K_Moffload4 tok/s$6995.3
NVIDIA GeForce RTX 5070Q4_K_Moffload4 tok/s$7505.1
NVIDIA GeForce RTX 4060 Ti 16GBQ4_K_Moffload4 tok/s$7705.3
AMD Radeon RX 7900 XTQ4_K_Moffload8 tok/s$8009.9
AMD Radeon RX 9070 XTQ4_K_Moffload5 tok/s$8156.0
NVIDIA GeForce RTX 4070 Ti SuperQ4_K_Moffload5 tok/s$9845.3
NVIDIA GeForce RTX 5070 TiQ4_K_Moffload5 tok/s$9955.4
NVIDIA GeForce RTX 4080 SuperQ4_K_Moffload5 tok/s$1,1994.4
AMD Radeon RX 7900 XTXQ4_K_Moffload21 tok/s$1,53313.6
NVIDIA GeForce RTX 5080Q4_K_Moffload6 tok/s$1,6003.4

2 other cards and machines we price cannot hold it at this context.

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k@128k
Q4_K_Mgguf4.8 GB6.7 GB10.0 GB14.4 GB23.3 GB
Q5_K_Mgguf5.6 GB7.5 GB10.8 GB15.2 GB24.1 GB
Q8_0gguf8.4 GB10.2 GB13.5 GB18.0 GB26.9 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/mistralai/Ministral-3-8B-Reasoning-2512-GGUF:Q4_K_M
    • LM Studiolms get mistralai/Ministral-3-8B-Reasoning-2512-GGUF@q4_k_m
    • llama.cppllama-server -hf mistralai/Ministral-3-8B-Reasoning-2512-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/mistralai/Ministral-3-8B-Reasoning-2512-GGUF:Q5_K_M
    • LM Studiolms get mistralai/Ministral-3-8B-Reasoning-2512-GGUF@q5_k_m
    • llama.cppllama-server -hf mistralai/Ministral-3-8B-Reasoning-2512-GGUF:Q5_K_M
  • Q8_0
    • Ollamaollama run hf.co/ggml-org/Ministral-3-8B-Reasoning-2512-GGUF:Q8_0
    • LM Studiolms get ggml-org/Ministral-3-8B-Reasoning-2512-GGUF@q8_0
    • llama.cppllama-server -hf ggml-org/Ministral-3-8B-Reasoning-2512-GGUF:Q8_0

Notes

Needs vllm >= 0.12.0 (with mistral_common >= 1.8.6) or Transformers 5.0.0rc0. Fits in 24GB of VRAM in BF16, or under 12GB when quantized.

Questions

How much VRAM does Ministral 3 8B Reasoning 2512 need?
At Q4_K_M, Ministral 3 8B Reasoning 2512 needs about 7 GB of memory at 8k context and 23 GB at 128k, counting the weights, the KV cache and runtime buffers.
What is the best GPU for running Ministral 3 8B Reasoning 2512 locally?
The cheapest hardware in the US that runs Ministral 3 8B Reasoning 2512 fully in memory at 128k context is the NVIDIA GeForce RTX 3090, $1,600 at eBay, at about 25 tok/s (estimated). The fastest we price is the NVIDIA GeForce RTX 5090, at about 41 tok/s.
Can I run Ministral 3 8B Reasoning 2512 on a 16 GB GPU?
Only partly: at 128k context it does not fit in 16 GB, so part of it runs from system RAM and it slows to about 5 tok/s on the Intel Arc A770 16GB (estimated).
Can I run Ministral 3 8B Reasoning 2512 on a 24 GB GPU?
Yes: at Q4_K_M it fits fully in 24 GB at 128k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 27 tok/s (estimated).
How do I run Ministral 3 8B Reasoning 2512 locally?
With Ollama: ollama run hf.co/mistralai/Ministral-3-8B-Reasoning-2512-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is Ministral 3 8B Reasoning 2512 released under?
Apache 2.0, which allows commercial use.