Arcee-Blitz: A Fast and Efficient 24B Mistral-Based Model

Arcee-Blitz is a 24 billion parameter model built on the Mistral architecture, developed by arcee-ai. It is distinguished by its distillation from DeepSeek-V3 logits, a process that imbues it with enhanced capabilities while maintaining efficiency. Designed as a practical "workhorse" model, Arcee-Blitz aims to handle a diverse range of tasks without the computational overhead typically associated with larger models.

Key Capabilities & Performance

• Improved World Knowledge: Arcee-Blitz shows substantial gains in world knowledge, reflected by a significant increase in MMLU-Pro scores (from 44.70% to 60.20%) compared to Mistral-Small-3.
• Enhanced Reasoning: Demonstrates better performance on reasoning tasks, with improvements in BBH and GPQA benchmarks.
• Strong Coding Performance: Outperforms Mistral-Small-3 across various BigCodeBench metrics, including Complete, Instruct, and Complete-hard categories.
• Mathematical Proficiency: Achieves a notable increase in Math Level 5 scores (from 12.00% to 38.60%).
• Efficient Architecture: Leverages a Mistral-based architecture, optimized for speed and resource efficiency.

Use Cases & Considerations

Arcee-Blitz is well-suited for applications requiring a capable yet efficient language model. Its improved world knowledge and coding abilities make it versatile for tasks such as content generation, code assistance, and complex question answering. The model has a context length of 32,768 tokens and is released under the Apache-2.0 License, allowing for broad commercial and non-commercial use. Users should be aware of its knowledge cut-off around June 2024 and potential risks of generating biased content, common to all large language models.