At a glance: Magistral Small, released June 10, 2025. 24B parameters. Apache 2.0. First open-weight reasoning model from Mistral. AIME 2024: 70.68%. GPQA Diamond: 68.18%. Fits an RTX 4090 at Q4. Ollama: magistral. Part of our AI Models & Companies reviews.
Magistral Small is the release that answered a question the open-weight AI community had been asking since DeepSeek R1 and QwQ-32B demonstrated that chain-of-thought reasoning could be distilled into smaller, local models: when would Mistral AI enter the reasoning model space?
The answer arrived on June 10, 2025, in the form of two models: Magistral Small (24B, Apache 2.0, open-weight) and Magistral Medium (proprietary, API-only). The Small variant is the one that matters for open deployment — the first Mistral model explicitly designed for multi-step logical reasoning, released under a license that permits commercial use and local deployment without restriction.
The model is built on Mistral Small 3.1 — inheriting its 24B dense architecture, 128K context capability, Tekken tokenizer, and multimodal vision. What distinguishes Magistral Small is its training: supervised fine-tuning on reasoning traces generated by Magistral Medium, followed by reinforcement learning using verifiable rewards (RLVR) — a technique that optimizes for correct final answers on problems with objective ground truth, not just plausible-sounding outputs.
The technical paper, arXiv:2506.10910, published June 12, 2025, describes the RLVR framework in detail and positions Magistral as Mistral’s first system in the reasoning model tier — a family alongside R1, o1, and QwQ rather than alongside generic instruction-tuned models.
Background: Why Reasoning Models in June 2025
By the time Magistral Small shipped, the reasoning model landscape had been defined by models that were either enormous (DeepSeek R1 at 671B total parameters), proprietary (OpenAI o1, o3), or trained on English-centric data (QwQ-32B, primarily English and Chinese).
The gap Magistral targets is specific: an open-weight reasoning model that is genuinely multilingual, fits on consumer GPU hardware under Apache 2.0, and does not require a closed API to access its chain-of-thought outputs. The chain-of-thought reasoning in Magistral is intended to appear in the user’s own language — not just output in the target language, but reason in it — which was rare in the June 2025 reasoning model cohort.
For context, here is the reasoning model landscape at the time of Magistral Small’s release:
| Model | Parameters | License | AIME 2024 | Hardware |
|---|---|---|---|---|
| DeepSeek R1 | 671B (MoE) | MIT | ~79.8% | Multi-node enterprise |
| QwQ-32B | 32B dense | Apache 2.0 | ~79.5% | 2× A100 or 80 GB GPU |
| Magistral Small | 24B dense | Apache 2.0 | 70.68% | Single RTX 4090 (Q4) |
| OpenAI o1 | Unknown | Proprietary | ~83.3% | API only |
Magistral Small does not top the AIME leaderboard. What it does is bring Apache 2.0 reasoning into the weight range that fits a single consumer GPU — something QwQ-32B cannot do at Q4 and R1 cannot approach without enterprise infrastructure.
Architecture
Magistral Small is architecturally identical to Mistral Small 3.1 at the base layer. It inherits the dense transformer design that Mistral has used across its Small series:
| Property | Value |
|---|---|
| Parameters | 24 billion |
| Base model | Mistral Small 3.1 (2503) |
| Context window | 40K optimal / 128K max |
| Architecture | Dense decoder-only Transformer |
| Attention | Grouped Query Attention (GQA) |
| Tokenizer | Tekken (131K vocab, byte-fallback BPE, Tiktoken-based) |
| Vision | Yes — multimodal (image + text input) |
| License | Apache 2.0 |
| HuggingFace | mistralai/Magistral-Small-2506 |
The Tekken tokenizer — introduced with Mistral NeMo in July 2024 and propagated through the entire Mistral Small 3.x series — brings the same compression advantages to Magistral Small: approximately 30% more efficient for European languages and source code versus the V3 tokenizer used in Mistral 7B and Mixtral, and significantly more efficient for Korean and Arabic. This matters for reasoning: longer token sequences mean higher inference costs and slower chain-of-thought generation. Better compression at the tokenizer level means the reasoning traces themselves are cheaper to generate.
The reasoning capability is a training property, not an architectural one. The base Mistral Small 3.1 architecture is unmodified; what changes is what the model learned to do:
- Supervised Fine-Tuning (SFT) on reasoning traces generated by Magistral Medium — a proprietary larger model. This teaches the 24B model to produce structured chain-of-thought outputs similar to those from a larger system.
- Reinforcement Learning with Verifiable Rewards (RLVR) applied on top of SFT. The model is rewarded for correct final answers on math, coding, and science problems where correctness can be verified objectively. This stage improves MATH and GPQA performance beyond the distilled baseline, though the paper notes it slightly reduces performance on code benchmarks compared to pure distillation.
The RLVR training approach is documented in arXiv:2506.10910. The combination of SFT and RL is what Mistral reports produces the best overall benchmark performance — neither approach alone achieves the same results.
Benchmarks
Magistral Small’s official benchmark results at release:
| Benchmark | Score | What It Tests |
|---|---|---|
| AIME 2024 pass@1 | 70.68% | Competition mathematics |
| LiveCodeBench v5 | 55.84% | Competitive programming |
| GPQA Diamond | 68.18% | Graduate-level science (PhD-level Q&A) |
Benchmark Context
AIME 2024 (70.68%): The American Invitational Mathematics Examination is the standard benchmark for reasoning model math performance. Magistral Small’s 70.68% is competitive but trails the leaders: QwQ-32B at approximately 79.5%, DeepSeek R1 at 79.8%, and OpenAI o1 at approximately 83.3%. The 8–9 point gap versus QwQ-32B is real — it reflects both the parameter advantage QwQ holds (32B vs 24B) and possibly the difference in reasoning-specific training depth.
LiveCodeBench v5 (55.84%): Competitive programming benchmark covering recent contest problems. QwQ-32B scores approximately 63.4%, and DeepSeek R1 approximately 65.9%. Magistral Small trails by ~8–10 points on code tasks, consistent with the AIME gap. The gap partially reflects the RLVR training tradeoff: Mistral’s paper notes that RL training on math and science slightly reduced LiveCodeBench performance compared to pure SFT distillation.
GPQA Diamond (68.18%): Graduate-level science questions requiring domain expertise in physics, chemistry, and biology. This is the benchmark where Magistral Small holds up most strongly relative to peers. GPQA tests a form of reasoning that benefits from multilingual scientific training, and Mistral’s 25-language training corpus may provide some advantage here.
Multilingual Reasoning
The single most differentiated claim for Magistral Small at launch is its multilingual chain-of-thought capability. Most reasoning models in June 2025 reason in English or Chinese regardless of input language — the chain-of-thought traces appear in the dominant language of the training data, and the model then translates or switches language for the final output.
Magistral Small is trained to produce reasoning traces in the user’s language. The model supports 25+ languages:
English, French, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Malay, Nepali, Polish, Portuguese, Romanian, Russian, Serbian, Spanish, Swedish, Turkish, Ukrainian, Vietnamese, Arabic, Bengali, Chinese (Simplified), Farsi.
The practical significance: a French-speaking researcher working through a mathematical proof, or a Japanese developer debugging code reasoning, can see the chain-of-thought in their own language rather than processing English-intermediate outputs. This is not a trivial capability — producing coherent logical chains in languages with different syntactic and mathematical notation conventions requires specific training, not just translation.
Hardware and Local Deployment
Magistral Small is the first open-weight reasoning model that fits on a single RTX 4090 (24 GB VRAM) at Q4 quantization. This is a meaningful threshold: the RTX 4090 is the highest-end consumer GPU available in 2025, and fitting within 24 GB at Q4 means local reasoning deployment is possible without enterprise hardware.
| Quantization | VRAM Required | Notes |
|---|---|---|
| BF16 (full precision) | ~48 GB | Requires 2× RTX 3090/4090 or single A100 |
| Q8_0 | ~25 GB | Slightly above RTX 4090 (24 GB); fits A6000 (48 GB) |
| Q4_K_M | ~14 GB | Single RTX 4090; also 16 GB gaming GPUs |
For Apple Silicon users: a 32 GB MacBook Pro M3/M4 can run Q4_K_M with headroom. A 64 GB MacBook Pro handles Q8_0. The unified memory architecture of Apple Silicon generally allows slightly higher effective utilization than discrete GPU VRAM.
Ollama: magistral
Available variants:
magistral:24b-small-2506-q4_K_M— approximately 14 GB, recommended for RTX 4090magistral:24b-small-2506-q8_0— approximately 25 GB, for A6000/H100 class hardware
HuggingFace: mistralai/Magistral-Small-2506
September 2025 update: A revised checkpoint, Magistral-Small-2509, was released in September 2025 with iterative improvements. Available at mistralai/Magistral-Small-2509 on HuggingFace.
Magistral Medium: The Proprietary Companion
Magistral Small was released alongside Magistral Medium, a larger proprietary model available via Mistral’s Le Chat platform and API. Magistral Medium is not open-weight — no downloadable weights, no local deployment.
The relationship between the two models is structural: Magistral Small was partly trained on reasoning traces from Magistral Medium (the SFT stage). Medium serves as the teacher, Small as the distilled and RL-refined student. Mistral claims Le Chat delivers Magistral responses at “10× the speed of competitors” — an inference efficiency claim for their hosted deployment, not a reflection of local Magistral Small throughput.
For users who need reasoning capability but do not require local deployment, Magistral Medium is accessible through the standard Mistral API under the model identifier magistral-medium-2506. For users who need open weights, local control, or Apache 2.0 rights, Magistral Small is the option.
Comparison to Contemporaries
vs. QwQ-32B (Alibaba/Qwen, March 2025)
QwQ-32B has 32B parameters versus Magistral Small’s 24B, scores higher on AIME (79.5% vs 70.68%) and LiveCodeBench (63.4% vs 55.84%), and also carries Apache 2.0. However, QwQ-32B requires approximately 20+ GB at Q4, approaching or exceeding RTX 4090 limits, and is primarily English/Chinese focused — not the 25-language reasoning coverage Magistral offers. If English-only math and code are the primary use case, QwQ-32B leads. For multilingual or hardware-constrained deployment, Magistral Small has the advantage.
vs. DeepSeek R1 (DeepSeek, January 2025)
DeepSeek R1 is 671B total parameters (MoE with ~37B active). AIME 79.8%, LiveCodeBench 65.9%. It requires multi-GPU enterprise hardware for full-precision inference and is MIT-licensed. The comparison is largely academic for local deployment — R1 simply cannot run on a single consumer GPU. Magistral Small provides 88% of R1’s AIME score at roughly 3.6% of the total parameter count, on accessible hardware.
vs. Mistral Small 4 (March 2026)
Mistral Small 4 (released March 2026) consolidates Magistral’s reasoning capability, Pixtral’s vision, and Devstral’s coding into a single 119B MoE model with only 6.5B active parameters per token. GPQA Diamond 71.2% versus Magistral Small’s 68.18%. Small 4 is the current recommended option for new deployments requiring reasoning — but Magistral Small remains relevant for deployments that need a smaller, simpler model without MoE routing complexity.
Strengths and Weaknesses
Strengths:
- First open-weight reasoning model from Mistral — marks entry into a new capability tier for Mistral’s open-source product line
- Apache 2.0 — permissive commercial use, no restrictions
- Single RTX 4090 deployment — the first Apache 2.0 reasoning model to hit this hardware target
- Multilingual chain-of-thought in 25+ languages — differentiated from English-centric peers
- Vision capability inherited from Mistral Small 3.1 — reasoning over images is possible
- Tekken tokenizer — efficient compression for European languages, code, and non-Latin scripts
Weaknesses:
- AIME 70.68% trails QwQ-32B (79.5%) and DeepSeek R1 (79.8%) — not the benchmark leader at launch or since
- Context ceiling: official optimal window is 40K tokens, not the full 128K of the base model — long-document reasoning degrades beyond this
- No dedicated code training: LiveCodeBench 55.84% lags behind specialized code reasoning models; the RLVR math/science focus slightly reduced code benchmark performance
- Verbose reasoning traces: chain-of-thought models produce long intermediate outputs — token budgets and latency are higher than standard instruction models
- Superseded: Mistral Small 4 (March 2026) provides improved reasoning at similar or better efficiency with MoE architecture — Magistral Small is the better choice only for users who prefer a simpler dense architecture or need a smaller active parameter count
Who Should Use Magistral Small
Strong fit:
- Developers building local reasoning pipelines who need Apache 2.0 and a single-GPU target
- Non-English language research, tutoring, or professional reasoning applications (French, German, Arabic, Japanese, etc.)
- Organizations evaluating reasoning model capabilities before committing to API-dependent closed models
- Applications where seeing the full chain-of-thought reasoning trace (in the user’s language) adds value — tutoring, explanation, audit trails
Better alternatives exist for:
- Highest AIME/MATH performance: QwQ-32B or DeepSeek R1
- Production-ready reasoning with more recent training: Mistral Small 4 (March 2026)
- API-only deployments at scale: Magistral Medium or OpenAI o3-mini
Rating: 4/5
Magistral Small earns its rating as the first meaningful reasoning model in Mistral’s open-weight lineup. The technical contribution is genuine: Apache 2.0, consumer GPU deployment, multilingual chain-of-thought, and RLVR training that Mistral has documented publicly. AIME 70.68% is real reasoning capability — it is not a trivial score — even though it trails QwQ-32B and DeepSeek R1.
The deduction from a perfect score reflects two real limitations: the AIME gap versus the reasoning model leaders, and the fact that Mistral Small 4 (less than a year later) consolidated reasoning, vision, and code into a more capable and efficient MoE architecture. Magistral Small is a genuine technical milestone that aged faster than its reach — which is a reasonable description of a fast-moving field, not a failure of the model itself.
For the historical record: in June 2025, Magistral Small was the most hardware-accessible Apache 2.0 reasoning model available. That mattered, and it earned the release.
This review is part of ChatForest’s ongoing coverage of the AI model landscape. See our Mistral AI company overview for context on Mistral’s full model lineup and our AI model reviews index for comparisons across other providers.