Darkhn/Magistral-2509-24B-Text-Only
Darkhn/Magistral-2509-24B-Text-Only is a 24 billion parameter, text-only causal language model derived from Mistral AI's Magistral Small 1.2, built upon Mistral Small 3.2 (2506). It was created by removing vision components to streamline performance for text-based tasks, retaining the original model's strong reasoning capabilities. This model excels at complex problem-solving and instruction following across dozens of languages, offering a 32K context window with good performance up to 40K tokens. It is optimized for applications requiring advanced text understanding and generation without multimodal overhead.
Loading preview...
Model Overview
Darkhn/Magistral-2509-24B-Text-Only is a 24 billion parameter, text-only model based on Mistral AI's Magistral Small 1.2, which itself is built on Mistral Small 3.2 (2506). This version was created by removing the vision encoder and multimodal layers, resulting in a more efficient model for purely text-based applications. It inherits the strong reasoning capabilities from its progenitor, having undergone Supervised Fine-Tuning (SFT) from Magistral Medium traces and subsequent Reinforcement Learning (RL).
Key Capabilities
- Advanced Reasoning: Capable of generating detailed, coherent chains of thought to solve complex problems.
- Multilingual Support: Proficient in dozens of languages, including English, French, German, Spanish, Italian, Japanese, Korean, Chinese, and Arabic.
- Optimized for Text: Streamlined for text-only tasks, offering a smaller footprint and potentially faster inference compared to its multimodal base.
- Extensive Context Window: Features a 32K token context window, with good performance maintained up to 40K tokens.
- Permissive Licensing: Released under the Apache 2.0 License, allowing for broad commercial and non-commercial use.
Reasoning Format
This model utilizes a specific reasoning format. The official format uses [THINK] and [/THINK] tags, requiring llama.cpp with --special and --jinja flags, and the /think keyword in the system prompt. A legacy format using <think> and </think> tags is effective with backends like Kobold.cpp and TabbyAPI, leveraging the base model's pre-training. Users should configure their system prompts and prefill sequences accordingly for optimal performance.
Benchmarks
Performance on text-based reasoning benchmarks is expected to be identical to the original Magistral Small 1.2, which shows strong results on AIME24 (86.14%), AIME25 (77.34%), GPQA Diamond (70.07%), and Livecodebench (70.88%).
Intended Use & Limitations
This model is ideal for text-based tasks requiring strong reasoning and instruction following. It is a text-only model and cannot process image inputs. Users should be aware of potential hallucinations and follow responsible AI practices, as safety alignment beyond the base model has not been added.