Ministral-3B-2410

Ministral-3B-2410 is explicitly documented as a dense Transformer model derived from Mistral Small 3.1 through a 'Cascade Distillation' process. Mistral AI provides technical details including the use of Grouped Query Attention (GQA), Rotary Position Embedding (RoPE), and RMS Normalization. The transition from the original 'les Ministraux' (October 2024) to the 'Ministral 3' series (December 2025) is documented, noting the addition of a 410M parameter vision encoder and the use of tied input-output embeddings to maintain the 3B scale. However, specific layer-by-layer architectural modifications for the 2410 variant specifically are less detailed than the later 2512 release.

Information regarding the training data is limited to high-level descriptions. Official documentation mentions a 'large proportion of multilingual and code data' and the use of 'publicly available and synthetic datasets.' While the 'Cascade Distillation' methodology explains how knowledge is transferred from the parent model (Mistral Small), there is no public breakdown of the specific data mixture, token counts per category, or detailed filtering/cleaning procedures for the 2410 version.

The model uses the V3-Tekken tokenizer, which is well-documented and publicly available via the 'mistral-common' library. It features a vocabulary size of 131,072 tokens and is based on tiktoken, moving away from the previous sentencepiece-based versions. The tokenizer's behavior, including the absence of leading whitespaces in chat templates, is explicitly detailed in the Mistral AI Cookbook and technical documentation.

The model is clearly identified as a dense architecture with approximately 3 billion parameters. For the 2410 variant, the parameter count is consistently reported. In the subsequent Ministral 3 (2512) update, the breakdown is even more transparent, specifying 3.4B for the language decoder and 410M for the vision encoder. The use of tied embeddings is also disclosed as a method to manage parameter density.

Mistral AI discloses that the models were trained on NVIDIA Hopper GPUs (H100/H200) through collaborations with NVIDIA. However, specific compute metrics such as total GPU hours, energy consumption, or carbon footprint for the Ministral-3B-2410 training run are not publicly available. The documentation focuses more on the efficiency of the distillation method rather than the absolute resources consumed.

Mistral provides standard benchmark results (MMLU, GSM8K, HumanEval) and compares them against competitors like Llama 3.2 3B and Gemma 2 2B. While evaluation code and specific prompt templates are not fully public in a single repository, third-party platforms like OpenRouter and Artificial Analysis have verified performance, though they occasionally report lower scores than official claims. The lack of a comprehensive technical paper for the 2410 version specifically (unlike the 2026 'Ministral 3' paper) limits full reproducibility.

The model consistently identifies itself as a Mistral-developed product. It maintains a clear versioning string (2410) and does not exhibit identity confusion with models from other providers. Documentation clearly distinguishes between the base, instruct, and later reasoning variants, ensuring users know which model they are interacting with.

The licensing for Ministral-3B-2410 is somewhat complex. While Mistral AI is known for open-source contributions, this specific model was initially released under the 'Mistral Commercial License' for self-deployment, with weights not initially available to researchers (unlike the 8B version). The later 'Ministral 3' (2512) update moved to Apache 2.0, but the 2410 variant remains under more restrictive commercial terms for many users, creating some ambiguity in the 'open' status of the weights.

Hardware requirements are well-documented for various deployment scenarios. Mistral and partners like NVIDIA and DigitalOcean provide specific VRAM guidance (e.g., ~8GB for 3B models). The model is optimized for vLLM and llama.cpp, with quantization support (FP8, INT4) and its impact on memory and latency being publicly discussed in technical blogs and community documentation.

Mistral uses date-based versioning (2410) and maintains a public changelog. However, there have been reports of silent updates, such as the November 2024 adjustment where the temperature parameter was downscaled by a multiplier of 0.43 across the Ministral family to 'unify model behavior.' While documented in the changelog, such changes can cause behavior drift for users relying on previous settings.