SWE-bench Verified
Qwen3.5-9B improved Pass@1 on SWE-bench Verified from 41.8% to 56.4%. Google Gemini 3.8 Flash solved 68% of the problems on the TheoPhysEval benchmark Google provides the Google Gemini 3.8 Flash model Open SWE -> SUPPORTS_PROVIDER -> Runloop The Qwen/Qwen3.5‑9B model is provided for use with the Transformers library. Flint leads coding benchmark SWE-Bench Verified Claude Code harness was tested on SWE-bench Verified OpenCode harness was tested on SWE-bench Verified mini-SWE-agent harness was tested on SWE-bench Verified· Graphiti 生成的实体摘要
关系变化
按生效时间- mini-SWE-agentTESTED ONSWE-bench Verified
mini-SWE-agent harness was tested on SWE-bench Verified
- Claude CodeTESTED ONSWE-bench Verified
Claude Code harness was tested on SWE-bench Verified
- OpenCodeTESTED ONSWE-bench Verified
OpenCode harness was tested on SWE-bench Verified
- FlintLEADSSWE-bench Verified
Flint leads coding benchmark SWE-Bench Verified
- SWE-bench VerifiedIS MODEL FORTransformers
The Qwen/Qwen3.5‑9B model is provided for use with the Transformers library.
- GooglePROVIDESSWE-bench Verified
Google provides the Google Gemini 3.8 Flash model
- SWE-bench VerifiedSOLVEDTheoPhysEval benchmark
Google Gemini 3.8 Flash solved 68% of the problems on the TheoPhysEval benchmark
- Open SWESUPPORTS PROVIDERSWE-bench Verified
Open SWE -> SUPPORTS_PROVIDER -> Runloop
- Qwen3.5-9BINCREASED PASS AT 1 ON SWE BENCH VERIFIED FROM 41.8 TO 56.4SWE-bench Verified
Qwen3.5-9B improved Pass@1 on SWE-bench Verified from 41.8% to 56.4%.
关系是 Graphiti 从报道中提取的事实陈述,不是原文引述;失效时间表示后来的报道显示这条关系已不成立。
相关事件 5
来源报道 6
- Fancy harnesses make coding agents more expensive but not better, study findsThe Decoder · 2026-10-24
- Recursive发布Flint语言模型The Decoder · 2026-10-22
- Hugging Face ❤️ TorchTPU:TPU 原生支持Hugging Face Blog · 一手 · 2026-10-20
- GPT-6-Astra 在理论物理基准测试中解决 49 道物理研究难题,成本仅 2.40 美元The Decoder · 2026-10-15
- Open SWE:一个用于内部编码代理的开源框架LangChain Blog · 2026-10-14
- Agent Lightning v1.0:3,500 行轻量化 Agentic RL 框架,支持真实 harness 训练Microsoft Research · 一手 · 2026-10-08