参考资料 6可信度 85%总体可信度: 85%该图钉的来源和参考资料对其日期的支持程度。包括来源在内的 6 份资料对图钉开始和结束时间支持程度的加权平均；资料每比最新的一份旧 180 天，权重减半显示所有可信度不低于 75% 的图钉
第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均；资料每比最新的一份旧 180 天，权重减半。
- [1]90%huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507huggingface.co· 发布于 2026年10月12日· 开始 2025年7月21日· 占评分 28%
Qwen 的 Qwen3 代码仓库更新日志写道：“2025.07.21：我们发布了 Qwen3-235B-A22B 非思考模式的更新版本，命名为 Qwen3-235B-A22B-Instruct-2507”；Simon Willison[5] 7 月 22 日的文章称其“昨天未作太多宣传就发布了”，OpenRouter[6] 则标注其发布于 Jul 21, 2025。
- [2]92%QwenLM/Qwen3 - GitHubgithub.com· 添加于 2026年10月12日· 开始 2025年7月21日 ✓· 占评分 28%
Qwen's repository news log: "2025.07.21: We released the updated version of Qwen3-235B-A22B non-thinking mode, named Qwen3-235B-A22B-Instruct-2507", then the Thinking-2507 sibling on 25 July and the 30B-A3B pair on 30 and 31 July; the 8 August update enables 1 million-token inputs.[1]
- [3]75%Qwen3 235B 2507 Instruct - Artificial Analysisartificialanalysis.ai· 添加于 2026年10月12日· 占评分 28%
Artificial Analysis: open-weights Qwen3 235B A22B 2507 Instruct, released July 2025, 256k context, scores 12 on its current Intelligence Index (median for comparable models 12), $0.22 per 1M input and $0.88 per 1M output tokens, now flagged deprecated in favour of Qwen3.5 397B A17B.
- [4]85%Alibaba's new open source Qwen3-235B-A22B-2507 beats Kimi-2 and offers low compute version - VentureBeatventurebeat.com· 发表于 2025年7月22日· 占评分 5%
VentureBeat, 22 July 2025: an instruct (non-reasoning) model plus an FP8 build; Qwen said it would "stop using hybrid thinking mode"; MMLU-Pro up from 75.2 to 83.0 and LiveCodeBench from 32.9 to 51.8; said to outperform Claude Opus 4 non-thinking and edge Kimi K2 on community charts; BF16 needs about 640GB of GPU memory, FP8 about 320GB.
- [5]85%Qwen/Qwen3-235B-A22B-Instruct-2507 - Simon Willison, 22 July 2025simonwillison.net· 发表于 2025年7月22日· 占评分 5%
Simon Willison: "Significant new model release from Qwen, published yesterday without much fanfare"; it ditches April's /no_think toggle and is exclusively non-reasoning; Apache 2.0; BF16 weights are 437.91GB and the FP8 variant 220.20GB; Qwen's own numbers have it beating Claude 4 Opus non-thinking on several tests, with no independent results yet.