AI Briefing
KO

Darwin-36B-Opus, GPQA 88.4%

·2026.04.26 04:13

Key point

A 36B MoE model built with Darwin V7 achieved 88.4% on GPQA Diamond.

Details

Darwin-36B-Opus is a 36B MoE model built on Qwen3.6-35B-A3B, combined with a partner distilled from Claude Opus 4.6 reasoning data.

  • It has a total 36B / active 3B structure, and supports 262K context, bfloat16, and Apache 2.0.
  • On public benchmarks, it recorded GPQA Diamond 88.4%, claiming the best performance in the Darwin series.
  • The evaluation method is a 2-pass protocol: an initial greedy pass, followed by majority-of-8 stochastic retry with a tiebreaker for incorrect answers.
  • Overall results improved from 145/198(73.2%) → 175/198(88.4%), with some shards achieving a perfect 25/25.
  • The creators explained that Darwin V7's recombination process takes under 1 hour on a single GPU, with the final merge taking under 10 minutes.
llmhuggingfacebenchmarkgguf
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditOpen-Weight CLM Released as Jev Alternative, Agent Decision-Making 4x FasterAnthropic NewsAnthropic Releases Claude Opus 5.5 with 40% Cost Reduction and Record Alignment ScoresHacker NewsStep 5 Preview Released: 600B Parameter MoE Model Highlights Agentic Tasks and Financial Analysis PerformanceRedditMica v0.1 4B Released: Open-Source Decision Model for Agent Loops Trained for Under $30PyTorchKRvLLM Releases 'Decision 1.0', AI Models Specialized for Decision-MakingHacker NewsQwen3.5-based Small Decision Model 'Kev' Released with TypeSafe API Compatibility and Performance Metrics