List your company
ResearchFormerly WizWand

LLM Chat Evaluation on ArenaHard

49.2AccuracyOfficial Instruct Model
-1.749611.477724.70537.9323May 31, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Official Instruct ModelProtocol=Official Instruct, Backbone=SmolLM3-3B2026.0549.2
GLM-4.7 (Scaffold-only)Protocol=Proposed Method, Backbone=SmolLM3-3B2026.0517.32
GPT-5.4 (High)Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.0514.34
GLM-4.7 & ANDES (Ours)Protocol=Proposed Method, Backbone=SmolLM3-3B2026.0514.11
Opus-4.7 (xHigh)Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.0510.21
Opus-4.6 (1M)Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.059
Opus-4.6Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.057.78
GPT-5.2Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.056.49
Gemini-3.1-ProProtocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.056.42
MiniMax-M2.5Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.052.18
Base Model (SmolLM3-3B)Protocol=Zero-Shot, Backbone=SmolLM3-3B2026.050.42
Kimi-K2-ThinkingProtocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.050.42
MiniMax-M2.1Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.050.42
GPT-5.1-Codex-MaxProtocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.050.42
GLM-4.7 (OpenCode)Protocol=Proposed Method, Backbone=SmolLM3-3B2026.050.42
Sonnet-4.5Protocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.050.24
Qwen3-MaxProtocol=Agent-Post-Trained, Backbone=SmolLM3-3B2026.050.21