Atlas

Benchmarks

← All benchmarks

WritingBench

Chat & Writing · 2025-03-07

WritingBench evaluates generative writing on 1,239 queries spanning six domains and 100 subdomains, with query-specific criteria scored by a fine-tuned critic model on a 1-10 scale and averaged. Some releases print the same mean multiplied by ten; Atlas stores the native 1-10 scale.

Top models (higher is better)

ModelScore
Qwen3 235B A22B Thinking 25078.8
Qwen3 Next 80B A3B Instruct8.7
Qwen3 VL 235B A22B Thinking8.7
Kimi K2 Instruct8.6
Qwen3 VL 32B Thinking8.6
Qwen3 30B A3B Instruct 25078.6
Qwen3 Omni 30B A3B Thinking8.6
Qwen3-VL-235B-A22B-Instruct8.6
Qwen3 VL 8B Thinking8.6
o38.5
Qwen3 235B A22B Instruct 25078.5
Qwen3 VL 30B A3B Thinking8.5
Qwen3 30B A3B Thinking 25078.5
Qwen3-Next-80B-A3B-Thinking8.5
Qwen3 VL 4B Thinking8.4
Gemini 2.5 Flash8.4
Qwen3 4B Instruct 25078.3
Qwen3 4B Thinking 25078.3
DeepSeek-R1-05288.3
Gemini 2.5 Pro8.3
Qwen3-VL-8B-Instruct8.3
Qwen3 VL 32B Instruct8.3
Qwen3 Omni 30B A3B Instruct8.3
Qwen3 VL 30B A3B Instruct8.3
Qwen3 VL 4B Instruct8.3
Qwen3-235B-A22B8.0
Opus 47.9
QwQ-32B7.9
O4 Mini7.8
Qwen3 14B7.8
Qwen3 VL 2B Thinking7.8
DeepSeek-R17.7
Qwen3-30B-A3B7.7
o1 (2024-12-17)7.7
ChatGPT-4o Latest observed 2025-03-277.5
o3-mini7.5
DeepSeek-V3-03247.5
Qwen3 4B7.3
Qwen3 VL 2B Instruct7.3
Gemma 3 27B IT7.2
Loading Atlas data…