olaverse/MIST-1-70B

MIST-1-70B: A Merged Llama 3.1 Model for Advanced Reasoning

MIST-1-70B is a 70 billion parameter model developed by olaverse, part of their MIST model family. It is constructed by blending four top Llama 3.1 70B models using the DARE+TIES merge method, which prunes redundant weights and resolves conflicts to combine their best capabilities. This approach aims to deliver a production-ready model with enhanced performance across various domains.

Key Capabilities

• Strong Reasoning: Incorporates DeepSeek R1 distillation for advanced reasoning abilities.
• High Helpfulness: Built upon models excelling in helpfulness benchmarks, such as Nemotron.
• Coding Proficiency: Generates clean, documented, and production-ready code.
• Mathematical Problem Solving: Offers step-by-step, structured solutions for math problems.
• Multilingual Support: Capable of understanding and generating text in over 8 languages.
• Long Context Window: Features a substantial 128K token context window for processing extensive inputs.
• Unrestricted Instruction Following: Designed to follow instructions without excessive refusals.

Usage Notes

Users should always utilize the built-in tokenizer.apply_chat_template for prompt formatting, as the model's mixed training heritage (Llama 3.1 and ChatML parents) can lead to issues like <|im_end|> tokens appearing in output if prompts are hardcoded. The model supports bfloat16 (requiring 140GB VRAM) and 4-bit quantized (requiring 40GB VRAM) precision.