- Core technology and architecture bright spots
- Open source, locally deployable
- Apply recommended scenes
Mistral AI announced the launch of its first linguistic model focused on reasoning Magistral aimed at addressing the shortcomings of the current mainstream LLM in the following areas:
-
** Lack of depth of knowledge in the field** and poor performance in professional tasks (e.g. financial modelling, legal analysis).
-
** The reasoning is not transparent** and it is difficult for users to track the logic of the model ‘ s conclusions.
-
** Weak multilingual reasoning**, especially in non-English languages.
-
** Lack of chain-based thinking support (Chain-of-thought, COT)** which makes complex issues difficult to solve.
Magistral is trying to construct an AI that “** can think like humans” - with the ability to be structured, verifiable and progressively reasoned, while supporting multilingual use.
-
Open source version of the Magistral Small:24B parameter (Apache 2.0 protocol), which can be downloaded on Hugging Face.
- Magistral Media: A more powerful version of the business, accessible through Le Chat, La Plateforme API, Amazon SageMaker, will be on line in the future, IBM WatsonX, Azure AI and Google Cloud.
Most votes (majority voting @64) provide the most consistent answer through multiple generation versions, indicating the ability to balance diversity and consistency. The Medium model achieved an accuracy rate of 90 per cent, close to the level of the top closed-source model GPT-4 Turbo, particularly in professional assignments.
-
Distinguished in the context of ** complex logical tasks, cross-cutting questions and answers, and interpretible reasoning**;
-
Operationalization in ** professional scenes (law, science and technology, academia)**;
- In the congener open source model (20B-30B) ** is the one with the strongest reasoning**.
Core technology and architecture bright spots
1. Chain-of-Thought Native Support Magistral originally supports chain-of-thinking (CoT) and is able to ** automatically generate interpretable chains of thinking. This is particularly critical for complex tasks (e.g. legal reasoning, systematic planning). **Resoning-Oriented
-
The model has been fine-tuned specifically for “multistep logical reasoning”.
-
Support for the creation of the Inner Thinking Path (inner Monologue), which presents a complete reasoning trajectory.
-
The output format is
... to distinguish between drafts and summaries and to improve interpretability.
2. Multilingual reasoning The model is multilingual in its reasoning, and in particular optimizes the following languages:
English, French, German, Spanish, Italian
- Arabic, Russian, Chinese.
Key points: Not only is multi-language input/output supported, ** the chain of reasoning can also be naturally generated in the target language. **3. Flash Answers + Tink Mode (business version only) Through the “Le Chat” platform, Magistral Media can open:
-
Flash Answers: 10 times the speed of reasoning in mainstream competitions (e.g. ChatGPT).
-
Think Mode: Optimizing multi-round reasoning response time and efficiency.
-
Actual performance: Almost “seconds out” results in structured tasks (e.g. decision tree reasoning, logical proof, code planning);
-
At the same time, accuracy and logic were maintained.
4. Transparency and traceability of reasoning Each generation is accompanied by a clear “thinking trajectory” that can be tracked and audited.
Open source, locally deployable
-
Models have been opened up to support commercial uses.
-
Quantified to run on single RTX 4090 or MacBook M2(32GB).
-
Provide multiple platform support, including: vLLM (recommended)
-
llama.cpp version (GGUF)
-
Support tools such as mstudio, llama, unsloth, Axolotl
** Chat template (recommended)** The best reasoning effect is obtained by using the default system prompt: [SYSTEM_PROMPT] You should first show inner monologue and give a clear summary of the answer. Your answer must be in the same language as the user. Please use Markdown format for output.
Apply recommended scenes
