Mistral Small 3vsDevstral Small 2 24B
Complete comparison between Mistral Small 3 (24B parameters, Mistral AI) and Devstral Small 2 24B (24B, Mistral AI). VRAM required by quantization, tokens per second measured on 4 reference GPUs, use-case verdict, license, installation commands. All figures are calculated from catalog data—no copy-pasting between pages.
In brief
| Feature | Mistral Small 3 | Devstral Small 2 24B |
|---|---|---|
| Settings | 24B | 24B |
| Family | Mistral | Mistral |
| Author | Mistral AI | Mistral AI |
| Origin | FR | FR |
| License | Apache 2.0 | Apache 2.0 |
| Context | 32 768 tokens | 256,000 tokens |
| Output | January 2025 | December 2025 |
Memory footprint
Approximate VRAM required for inference with an average context window. The winner (in green) is the model that uses less — smaller models have the edge.
| Quantization | Mistral Small 3 | Devstral Small 2 24B |
|---|---|---|
| Q4_K_M (lightweight) | 14 GB | 14 GB |
| Q5_K_M (balanced) | 17 GB | 17 GB |
| Q8 (near-lossless) | 26 GB | 26 GB |
| FP16 (max quality) | 48 GB | 48 GB |
| CPU-only RAM | 24 GB | 24 GB |
Estimated speed (tokens/second)
Estimates based on the best quantization that can fit on each GPU. Actual figures depend on the prompt, context, and engine (llama.cpp, vLLM, MLX). Methodology.
| Reference GPU | Mistral Small 3 | Devstral Small 2 24B |
|---|---|---|
| RTX 4090 (24 GB) | 40 tok/s · Q5_K_M | 40 tok/s · Q5_K_M |
| RTX 4080 (16 GB) | 15 tok/s · Q4_K_M | 15 tok/s · Q4_K_M |
| RTX 3060 12GB (12 GB) | ✗ too heavy | ✗ too heavy |
| Apple M4 Pro (48 GB) (36 GB) | 15 tok/s · Q8 | 15 tok/s · Q8 |
Verdict by use case
For each common use case, we indicate which of the two is better suited based on its tags, size, and specialization.
Replace GitHub Copilot and Cursor with a 100% local coding copilot—the definitive guide, configs included.
- Lifetime online access
- PDF + files
- 30-day money-back guarantee
Strengths and weaknesses
Mistral Small 3
Excellent quality-to-size ratio at release (early 2025). Competes with 70B models.
- Quality close to Llama 70B
- Low latency
- 128k context
- Excellent for agents and tool use
- Requires at least 16 GB of VRAM in Q4
- Worse at coding than Qwen Coder
Installation
Devstral Small 2 24B
24B coding specialist, Apache 2.0. 72.2% SWE-Bench. 256k ctx, FR lab.
- 68.0% SWE-bench Verified
- Fits on RTX 4090
- Apache 2.0
- French lab
- No vision
- Dense: slower than code MoEs (Qwen3-Coder 30B-A3B) at the same memory capacity
Installation
Frequently asked questions
Which model runs best on RTX 4090 (24 GB): Mistral Small 3 or Devstral Small 2 24B?
On an RTX 4090, Mistral Small 3 runs at Q5_K_M (~40 tok/s), Devstral Small 2 24B in Q5_K_M (~40 tok/s). In pure throughput, Devstral Small 2 24B wins. See the configurator to test your exact GPU.
Mistral Small 3 or Devstral Small 2 24B: which is better for French?
For formal French (writing, summarization, translation), Mistral Small 3 has the edge—trained on a dense French-language corpus. The other remains usable, but for less demanding use cases.
Which uses the least VRAM?
Both require the same VRAM in Q4 (14 GB). The choice then comes down to the license or perceived quality.
Mistral Small 3 or Devstral Small 2 24B for coding?
Both are specialized code models — the choice comes down to context (32,768 vs 256,000 tokens), license (Apache 2.0 vs Apache 2.0), and available VRAM. Both integrate with Continue or Aider.
Can these models be used in commercial production?
Mistral Small 3 is under Apache 2.0 — free for commercial use. Devstral Small 2 24B is under Apache 2.0 — also free to use. For SaaS, prefer Apache 2.0 or MIT.
Which one should you choose in 2026?
Depends on your primary constraint. Smaller / faster: Mistral Small 3 (24B). More capable: Mistral Small 3 (24B). If you're unsure, run the configurator with your GPU and use case — it will decide based on both.
Alternatives to consider
If neither works for you, here are some neighboring models you could explore.