Devstral Small 2507
23.6B parameters · 128k context · Apache 2.0 · released 2025-07-04
codingagentic
In short
Devstral Small 2507 is an open model from Mistral AI with 23.6B parameters. At Q4_K_M, Devstral Small 2507 needs about 15 GB of memory at 8k context and 35 GB at 128k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in Canada that runs Devstral Small 2507 fully in memory at 128k context is the NVIDIA RTX PRO 6000 Blackwell, $3,133 at Amazon, at about 24 tok/s (estimated).
Updated . Speeds are estimates; how they are worked out.
What to run it on in Canada
Fully in memory at 128k context, priced at retailers today.
Cheapest
NVIDIA RTX PRO 6000 Blackwell24tok/sest.Q8_0
$3,133
Best value
NVIDIA RTX PRO 6000 Blackwell24tok/sest.Q8_0
$3,133
Fastest
NVIDIA RTX PRO 6000 Blackwell24tok/sest.Q8_0
$3,133
All hardware that runs it
Everything we price, best value first. Two cards split the model by layers.
| Hardware | Quant | Fits | Decode | Price | tok/s per 1,000 |
|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell | Q8_0 | fits | 24 tok/s | $3,133 | 7.6 |
| GMKtec EVO-X2 Ryzen AI Max+ 395 128GB | Q8_0 | fits | 3 tok/s | $5,099 | 0.6 |
| ASUS Ascent GX10 128GB | Q8_0 | fits | 4 tok/s | $8,299 | 0.4 |
| Mac Studio M4 Max 128GB | Q8_0 | fits | 8 tok/s | – | – |
| Mac mini M4 Pro 64GB | Q8_0 | fits | 4 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 128GB | Q8_0 | fits | 3 tok/s | – | – |
| Mac Studio M3 Ultra 512GB | Q8_0 | fits | 11 tok/s | – | – |
| NVIDIA DGX Spark 128GB | Q8_0 | fits | 4 tok/s | – | – |
| Mac Studio M3 Ultra 256GB | Q8_0 | fits | 11 tok/s | – | – |
| Mac mini M4 Pro 48GB | Q4_K_M | fits | 5 tok/s | – | – |
| Mac Studio M4 Max 64GB | Q8_0 | fits | 8 tok/s | – | – |
| Mac Studio M3 Ultra 96GB | Q8_0 | fits | 11 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 64GB | Q8_0 | fits | 3 tok/s | – | – |
| NVIDIA GeForce RTX 3060 12GB | Q4_K_M | offload | 2 tok/s | $497 | 3.8 |
| Intel Arc B580 | Q4_K_M | offload | 2 tok/s | $560 | 3.4 |
| Intel Arc A770 16GB | Q4_K_M | offload | 2 tok/s | $655 | 3.4 |
| NVIDIA GeForce RTX 3080 10GB | Q4_K_M | offload | 2 tok/s | $769 | 2.5 |
| AMD Radeon RX 9060 XT 16GB | Q4_K_M | offload | 2 tok/s | $800 | 2.6 |
| AMD Radeon RX 9070 XT | Q4_K_M | offload | 2 tok/s | $1,050 | 2.2 |
| Intel Arc Pro B60 | Q4_K_M | offload | 3 tok/s | $1,060 | 2.8 |
| NVIDIA GeForce RTX 5060 Ti 16GB | Q4_K_M | offload | 2 tok/s | $1,120 | 2.1 |
| NVIDIA GeForce RTX 5070 | Q4_K_M | offload | 2 tok/s | $1,181 | 1.7 |
| AMD Radeon RX 7800 XT | Q4_K_M | offload | 2 tok/s | $1,397 | 1.6 |
| NVIDIA GeForce RTX 4070 Super | Q4_K_M | offload | 2 tok/s | $1,400 | 1.4 |
| AMD Radeon RX 7600 XT | Q4_K_M | offload | 2 tok/s | $1,440 | 1.5 |
| NVIDIA GeForce RTX 5070 Ti | Q4_K_M | offload | 2 tok/s | $1,690 | 1.4 |
| NVIDIA GeForce RTX 4070 Ti Super | Q4_K_M | offload | 2 tok/s | $1,790 | 1.3 |
| NVIDIA GeForce RTX 4080 Super | Q4_K_M | offload | 2 tok/s | $2,000 | 1.2 |
| AMD Radeon RX 7900 XT | Q4_K_M | offload | 3 tok/s | $2,083 | 1.3 |
| NVIDIA GeForce RTX 5080 | Q4_K_M | offload | 2 tok/s | $2,370 | 1.0 |
| NVIDIA GeForce RTX 3090 | Q4_K_M | offload | 4 tok/s | $2,800 | 1.3 |
| AMD Radeon AI PRO R9700 | Q4_K_M | offload | 6 tok/s | $2,800 | 2.2 |
| NVIDIA GeForce RTX 3090 Ti | Q4_K_M | offload | 4 tok/s | $3,999 | 1.0 |
| NVIDIA GeForce RTX 4090 | Q4_K_M | offload | 4 tok/s | $7,107 | 0.5 |
| NVIDIA GeForce RTX 5090 | Q4_K_M | offload | 10 tok/s | $10,300 | 1.0 |
| AMD Radeon RX 7900 XTX | Q4_K_M | offload | 4 tok/s | – | – |
| NVIDIA GeForce RTX 4060 Ti 16GB | Q4_K_M | offload | 2 tok/s | – | – |
| NVIDIA Tesla P40 | Q4_K_M | offload | 3 tok/s | – | – |
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).
| Quant | File | @8k | @32k | @64k | @128k |
|---|---|---|---|---|---|
| Q4_K_Mgguf | 13.3 GB | 15.3 GB | 19.2 GB | 24.4 GB | 34.8 GB |
| Q5_K_Mgguf | 15.6 GB | 17.6 GB | 21.5 GB | 26.7 GB | 37.1 GB |
| Q8_0gguf | 23.3 GB | 25.3 GB | 29.2 GB | 34.4 GB | 44.8 GB |
Run it
Copy a command for the runtime you use.
- Q4_K_M
- Ollama
ollama run hf.co/mistralai/Devstral-Small-2507_gguf:Q4_K_M - LM Studio
lms get mistralai/Devstral-Small-2507_gguf@q4_k_m - llama.cpp
llama-server -hf mistralai/Devstral-Small-2507_gguf:Q4_K_M
- Ollama
- Q5_K_M
- Ollama
ollama run hf.co/mistralai/Devstral-Small-2507_gguf:Q5_K_M - LM Studio
lms get mistralai/Devstral-Small-2507_gguf@q5_k_m - llama.cpp
llama-server -hf mistralai/Devstral-Small-2507_gguf:Q5_K_M
- Ollama
- Q8_0
- Ollama
ollama run hf.co/mistralai/Devstral-Small-2507_gguf:Q8_0 - LM Studio
lms get mistralai/Devstral-Small-2507_gguf@q8_0 - llama.cpp
llama-server -hf mistralai/Devstral-Small-2507_gguf:Q8_0
- Ollama
Notes
Text-only 24B coding agent (vision encoder removed), said to run on a single RTX 4090 or 32GB Mac. vLLM serving requires vLLM >= 0.9.1 and mistral_common >= 1.7.0.
Questions
- How much VRAM does Devstral Small 2507 need?
- At Q4_K_M, Devstral Small 2507 needs about 15 GB of memory at 8k context and 35 GB at 128k, counting the weights, the KV cache and runtime buffers.
- What is the best GPU for running Devstral Small 2507 locally?
- The cheapest hardware in Canada that runs Devstral Small 2507 fully in memory at 128k context is the NVIDIA RTX PRO 6000 Blackwell, $3,133 at Amazon, at about 24 tok/s (estimated).
- Can I run Devstral Small 2507 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 2 tok/s on the Intel Arc A770 16GB (estimated).
- Can I run Devstral Small 2507 on a 24 GB GPU?
- Only partly: at 128k context it does not fit in 24 GB, so part of it runs from system RAM and it slows to about 3 tok/s on the Intel Arc Pro B60 (estimated).
- How do I run Devstral Small 2507 locally?
- With Ollama: ollama run hf.co/mistralai/Devstral-Small-2507_gguf:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
- What licence is Devstral Small 2507 released under?
- Apache 2.0, which allows commercial use.