| Property | Value |
|---|---|
| Parameter Count | 24B |
| Base Architecture | Mistral-Small-24B-Instruct-2501 |
| License | Apache-2.0 |
| Context Length | 32k Tokens |
| Model URL | https://huggingface.co/arcee-ai/Arcee-Blitz |
What is Arcee-Blitz?
Arcee-Blitz is a 24B parameter language model that represents a significant advancement in efficient AI model design. Built on the Mistral architecture and distilled from DeepSeek, it's designed to be a practical "workhorse" model that delivers robust performance across various tasks while maintaining computational efficiency.
Implementation Details
The model leverages a sophisticated distillation process, incorporating over 3B tokens of pretraining distillation from DeepSeek-V3 logits. The implementation includes a merged Virtuoso pipeline with the Mistral architecture, followed by additional fine-tuning steps to optimize performance.
- Advanced distillation process from DeepSeek-V3
- Merged Virtuoso pipeline integration
- Extensive post-training optimization
- Support for both GGUF and AWQ quantizations
Core Capabilities
- Significant improvements in MMLU-Pro performance (60.20% vs 44.70% baseline)
- Enhanced mathematical reasoning (Math Level 5: 38.60% vs 12.00% baseline)
- Improved world knowledge and general task performance
- Strong performance in code-related tasks (BigCodeBench improvements)
Frequently Asked Questions
Q: What makes this model unique?
Arcee-Blitz stands out for its efficient architecture that maintains high performance while being optimized for practical use. The model shows significant improvements in world knowledge and mathematical reasoning, making it particularly valuable for real-world applications.
Q: What are the recommended use cases?
The model is well-suited for a wide range of applications including code generation, mathematical problem-solving, and general language understanding tasks. Its 32k token context length makes it particularly useful for handling longer documents and complex queries.