Devstral-Small-2507 needs about 16.2 GB of memory at Q4_K_M (a 14.3 GB file) with an 8K context window, so it fits on a 24 GB GPU. This repository has 26 quants, from 5.6 GB (IQ1_S) to 47.2 GB (BF16). Longer context needs more: at 32K tokens the KV cache alone adds about 5.4 GB.
24B‑parameter Mistral model for agentic coding tasks
Devstral‑Small‑2507 is a 24 billion‑parameter LLM fine‑tuned for software engineering. It supports a 128k‑token context window and is text‑only, having removed the vision encoder. The model is released under Apache 2.0 and can run on a single RTX 4090 or a Mac with 32 GB RAM.
Good at
- Agentic coding with tool‑calling
- Long‑context code analysis
- Function‑calling compatible
Limits
- No vision support after fine‑tuning
- Requires 8‑k context for memory estimates
- Apache 2.0 license limits commercial use to open‑source terms
Best for
- Local coding assistants
- OpenHands integration
- On‑device software engineering agents
Who it’s for. Developers and teams needing a lightweight, agentic coding model that runs locally. The model is suitable for open‑source projects and commercial use under Apache 2.0.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| IQ1_S | 5.6 GB | 7.4 GB | Tight | Get file → | |
| IQ1_M | 6.0 GB | 7.9 GB | Tight | Get file → | |
| IQ2_XXS | 6.8 GB | 8.6 GB | Too big | Get file → | |
| IQ2_M | 8.2 GB | 10.1 GB | Too big | Get file → | |
| Q2_K | 8.9 GB | 10.7 GB | Too big | Get file → | |
| Q2_K_L | 9.0 GB | 10.9 GB | Too big | Get file → | |
| Q2_K_XL | 9.3 GB | 11.1 GB | Too big | Get file → | |
| IQ3_XXS | 9.4 GB | 11.3 GB | Too big | Get file → | |
| Q3_K_S | 10.4 GB | 12.2 GB | Too big | Get file → | |
| Q3_K_M | 11.5 GB | 13.3 GB | Too big | Get file → | |
| Q3_K_XL | 11.9 GB | 13.7 GB | Too big | Get file → | |
| IQ4_XS | 12.8 GB | 14.6 GB | Too big | Get file → | |
| IQ4_NL | 13.5 GB | 15.3 GB | Too big | Get file → | |
| Q4_0 | 13.5 GB | 15.3 GB | Too big | Get file → | |
| Q4_K_S | 13.5 GB | 15.4 GB | Too big | Get file → | |
| Q4_K_M | 14.3 GB | 16.2 GB | Too big | Get file → | |
| Q4_K_XL | 14.5 GB | 16.4 GB | Too big | Get file → | |
| Q4_1 | 14.9 GB | 16.7 GB | Too big | Get file → | |
| Q5_K_S | 16.3 GB | 18.1 GB | Too big | Get file → | |
| Q5_K_M | 16.8 GB | 18.6 GB | Too big | Get file → | |
| Q5_K_XL | 16.8 GB | 18.6 GB | Too big | Get file → | |
| Q6_K | 19.3 GB | 21.2 GB | Too big | Get file → | |
| Q6_K_XL | 20.8 GB | 22.6 GB | Too big | Get file → | |
| Q8_0 | 25.1 GB | 26.9 GB | Too big | Get file → | |
| Q8_K_XL | 29.0 GB | 30.8 GB | Too big | Get file → | |
| BF16 | 47.2 GB | 49.0 GB | Too big | Get file → |
For 8 GB VRAM use IQ1_S (5.56 GB, ~7.4 GB RAM). With 12 GB VRAM IQ4_K_M (14.33 GB, ~16.2 GB RAM) is a good balance. For 24 GB VRAM, Q6_K (19.35 GB, ~21.2 GB RAM) provides higher quality.
What GPU do you need for Devstral-Small-2507?
These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.
- 4 GB VRAM: no quant fits at an 8K context.
- 6 GB VRAM: no quant fits at an 8K context.
- 8 GB VRAM: the best quant that fits is IQ1_M (6.0 GB file, about 7.9 GB in use).
- 12 GB VRAM: the best quant that fits is IQ3_XXS (9.4 GB file, about 11.3 GB in use).
- 16 GB VRAM: the best quant that fits is Q4_K_S (13.5 GB file, about 15.4 GB in use).
- 24 GB VRAM: the best quant that fits is Q6_K_XL (20.8 GB file, about 22.6 GB in use).
- 32 GB VRAM: the best quant that fits is Q8_K_XL (29.0 GB file, about 30.8 GB in use).
- 48 GB VRAM: the best quant that fits is Q8_K_XL (29.0 GB file, about 30.8 GB in use).
Which Devstral-Small-2507 quant should you choose?
Q4_K_M is the usual starting point for Devstral-Small-2507: it balances file size and quality well. Here it is a 14.3 GB file.
The Q8_0 file is 25.1 GB, about 1.7 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.
If memory is tight, the smallest file is IQ1_S at 5.6 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about Devstral-Small-2507
How much VRAM does Devstral-Small-2507 need?
About 16.2 GB at Q4_K_M with an 8K context window. That is the 14.3 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.
Can I run Devstral-Small-2507 on an 8 GB GPU?
Yes, with IQ1_M: the 6.0 GB file needs about 7.9 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of Devstral-Small-2507 GGUF?
It depends on the quant: IQ1_S is 5.6 GB, IQ1_M is 6.0 GB, IQ2_XXS is 6.8 GB, IQ2_M is 8.2 GB, Q2_K is 8.9 GB, Q2_K_L is 9.0 GB and 20 more.
What context length does Devstral-Small-2507 support?
The GGUF header lists a trained context length of 131,072 tokens (128K). The KV cache grows with the context you use: about 1.3 GB at 8K and 5.4 GB at 32K.
Where can I download Devstral-Small-2507 GGUF?
The files are on Hugging Face in the unsloth/Devstral-Small-2507-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.
What license does Devstral-Small-2507 use?
Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.
Can I run Devstral‑Small‑2507 on an RTX 4090 with 8 GB VRAM?
Use the IQ1_S quant, which is 5.56 GB and requires about 7.4 GB RAM at 8k context.
Does this model support image input?
No, the vision encoder was removed during fine‑tuning; the model is text‑only.
What license applies to this model?
The model is released under Apache 2.0, allowing commercial and non‑commercial use with attribution.