Magistral Small

The model's lineage is clearly documented as being built upon Mistral Small 3.1 (2503). The technical architecture is well-defined as a 24B dense decoder-only transformer with 32 layers, utilizing Grouped Query Attention (GQA), SwiGLU activations, and Rotary Positional Embeddings (RoPE). The training methodology is explicitly described in the accompanying paper (arXiv:2506.10910), detailing a 'cold-start' SFT process using reasoning traces from Magistral Medium followed by Reinforcement Learning from Verifiable Rewards (RLVR).

While the training methodology (SFT traces + RLVR) is well-documented, the specific composition of the underlying pre-training data for the base model (Mistral Small 3.1) remains largely undisclosed beyond general categories. The reasoning-specific data is described as being derived from 'Magistral Medium traces' and 'verifiable reward' tasks (math, code), but precise dataset distributions, source names, and filtering/cleaning metrics for the vast majority of the training corpus are absent.

The model uses the 'Tekken' tokenizer, which is part of the publicly available 'mistral-common' library. It features a vocabulary size of 131,072 tokens and is specifically optimized for over 24 languages and source code. The tokenizer's performance and alignment with the model's multilingual claims are verifiable through the public GitHub repository and Hugging Face model files.

The model is explicitly stated to be a 24.0 billion parameter dense architecture. Unlike Mixture-of-Experts (MoE) models where active parameters can be ambiguous, this dense configuration ensures all 24B parameters are active during inference. Detailed architectural specifications, including the hidden dimension (14,336) and head counts (32 attention, 8 KV), are provided in the technical documentation.

Information regarding the specific compute resources used for training is minimal. While the paper discusses the 'asynchronous system' and infrastructure for online RL, it lacks concrete data on total GPU/TPU hours, hardware counts, carbon footprint, or energy consumption. The disclosure is limited to high-level descriptions of the training pipeline rather than quantifiable resource metrics.

Mistral provides specific results for standard benchmarks (AIME24, GPQA, LiveCodeBench) and includes the exact sampling parameters (top_p: 0.95, temp: 0.7) and system prompts required to replicate the reasoning behavior. Third-party reproduction guides (e.g., via promptfoo) are already available. However, the full evaluation code and the complete set of internal prompts used for all reported metrics are not entirely public.

The model demonstrates high identity consistency, correctly identifying itself as a Mistral-developed reasoning model. It maintains clear versioning (2506) and is transparent about its specific 'Think' mode capabilities and the 40k-128k context window limitations. There are no documented instances of the model claiming to be a competitor or misrepresenting its 24B scale.

The model is released under the highly permissive Apache 2.0 license, which is explicitly stated on Hugging Face, the official blog, and in the technical paper. This license allows for unrestricted commercial use, modification, and distribution, providing maximum legal clarity for downstream users without conflicting proprietary terms.

Hardware requirements are well-documented for various deployment scenarios. Official documentation specifies that the model fits on a single RTX 4090 (24GB VRAM) or a 32GB RAM Mac when quantized (4-bit). VRAM estimates for different quantization levels (Q4, Q8) are provided by the community and supported by official GGUF releases, with clear guidance on the 40k token context performance threshold.

Mistral uses a date-based semantic versioning system (2506) and maintains a clear distinction between variants (Small vs. Medium). While a changelog for the 'Magistral' family is emerging, the model is still relatively new, and long-term tracking of behavioral drift or silent updates is not yet fully established. Previous versions remain accessible on Hugging Face.