gm8xx8

DeepSeek v2.5 
🔗: https://huggingface.co/deepseek-ai/DeepSeek-V2.5
Combines DeepSeek v2 and DeepSeek v2 Coder with a 128K context and a 238B-parameter MoE architecture (21B active), optimized for coding.

“DeepSeek-V2.5 outperforms both DeepSeek-V2-0628 and DeepSeek-Coder-V2-0724 on most benchmarks.”

- HumanEval: 89% (py), 73.8% (all)
- Arena Hard: 76.2% (up from 68.3%)
- AlpacaEval: 50.5% (up from 46.6%)
- MT Bench: 9.02 (up from 8.84)
- LiveCodeBench: 41%
- Enhanced coding, writing, and instruction-following. 
- Supports function calling, fill-in-the-middle, and JSON output. 

banger—don’t sleep on the whale.