⚛️量子位•較早收集於 6h
千問 3.5 登頂 Hugging Face 排行榜
#open-source-llm#benchmark-win#moe#chinese-aiqwen-3.5qwen-3.5hugging-face
💡Qwen 3.5 #1 open model, China 8/10 top—stronger than Llama 3.1? Test it!
⚡ 30-Second TL;DR
有什麼變化
千問 3.5 榮登 Hugging Face 開源排行榜第一
為什麼重要
鞏固中國在開源權重 LLM 的主導地位，迫使西方實驗室加速發布。
下一步行動
Download Qwen-3.5 from Hugging Face and run it on your MMLU or coding benchmarks.
誰應關注：Developers & AI Engineers
關鍵要點
- •千問 3.5 榮登 Hugging Face 開源排行榜第一
- •中國模型佔前十的 8 席
- •總 3970 億參數，MoE 激活 170 億
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 6 個來源。
🔑 增強重點摘要
- •Qwen3-Next-80B ranks third on the F5 CASI security leaderboard with a score of 81.10, behind Claude Sonnet 4 and GPT-5[1].
- •Qwen3-30B-A3B appears on the F5 CASI leaderboard with a score of 58.33 in a lower tier, indicating varied performance across model sizes[1].
- •Qwen 3.5 is ranked in the A-tier on the Onyx AI Open Source LLM Leaderboard 2026, alongside MiMo-V2-Flash[6].
- •Qwen3-235B-A22B received a 0725 update but scored only 50.97 on the updated CASI benchmark, vulnerable to attack agents[1].
🔮 前景展望AI analysis grounded in cited sources
Chinese models will capture 9/10 top spots on Hugging Face open leaderboard by mid-2026
Current dominance of 8/10 positions combined with rapid iterations like Qwen3-235B-A22B updates signals aggressive advancement.
MoE efficiency in Qwen 3.5 will pressure closed models to release open variants
Activating only 170B of 397B parameters achieves top performance, challenging high-cost proprietary models on leaderboards like CASI.
📎 來源 (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： 量子位 ↗
每週 AI 簡報
每週一封，可隨時退訂。