🕷️
SetupAI
EN
🧠較早收集於 17m

阿里開源 3 款 Qwen3.5 模型，消費級顯卡可跑

阿里開源 3 款 Qwen3.5 模型，消費級顯卡可跑
PostLinkedIn
🧠閱讀原文： 机器之心
#open-source-llm#moe-architecture#consumer-gpuqwen3.5alibabaqwen3.5qwen3.5-27bqwen3.5-flash

💡New open-weight Qwen3.5 beats GPT-5 mini, runs on consumer GPUs—deploy locally now

⚡ 30-Second TL;DR

有什麼變化

開源 Qwen3.5-35B-A3B、Qwen3.5-122B-A10B、Qwen3.5-27B 模型

為什麼重要

實現頂級 LLM 低成本本地部署，降低開發者門檻，提升開源對閉源競爭。阿里雲廉價 API 加速企業採用。

下一步行動

Download Qwen3.5-27B from Hugging Face and test on your consumer GPU for agent tasks.

誰應關注：Developers & AI Engineers

關鍵要點

  • •開源 Qwen3.5-35B-A3B、Qwen3.5-122B-A10B、Qwen3.5-27B 模型
  • •在 GPQA、SWE-bench 等榜單超越 Qwen3-235B-A22B、Qwen3-VL 及 GPT-5 mini
  • •Qwen3.5-27B 密集模型單 GPU 運行，具強大 Agent 及多模態能力
  • •Qwen3.5-Flash API：1M 上下文，每百萬輸入 Token 0.2 元

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 9 個來源。

🔑 增強重點摘要

  • •Qwen3.5-397B-A17B achieves 19x faster decoding on long-context tasks (256k tokens) compared to Qwen3-Max while matching its reasoning and coding performance[7].
  • •Qwen3.5 excels in agentic terminal coding with 52.5 on Terminal-Bench 2.0, surpassing Qwen3-Max (22.5) and approaching Gemini 3 Pro (54.2)[7].
  • •In document recognition, Qwen3.5 scores 90.8 on OmniDocBench v1.5, outperforming GPT-5.2 (85.7) and Claude Opus 4.5 (87.7)[7].
📊 競品分析▸ Show
FeatureQwen3.5-Flash/Qwen3 VLGPT-5 Mini
Context Window262k tokens400k tokens[1][2][3]
Input Cost0.2 yuan (~$0.028)/M tokens (Flash); cheaper overall~$0.25/M tokens[2][5][6]
Output CostNot specified; generally lower~$2/M tokens[2][5][6]
Coding Benchmarks16.5 (Qwen3 VL); strong in Terminal-Bench35.3; leads in LiveCodeBench (83.8)[1][7]
Speed (tok/s)44.4 (Qwen3 VL)130.7[1]
Intelligence Index20.6 (Qwen3 VL)41.0[1]

🔮 前景展望AI analysis grounded in cited sources

Qwen3.5 series will capture >20% more open-source agentic AI deployments by mid-2026
Its superior Terminal-Bench performance and consumer GPU compatibility lower barriers for developers building autonomous agents compared to proprietary rivals[7].
Alibaba Cloud API pricing undercuts OpenAI by 80%+ on input tokens
At 0.2 yuan per million (~$0.028 USD), Qwen3.5-Flash offers massive cost savings over GPT-5 Mini's $0.25, driving high-volume enterprise adoption[1][2].

⏳ 時間線

2025-07
Qwen3 Coder 480B A35B released as specialized coding model
2025-08
OpenAI launches GPT-5 Mini with 400k context window
2025-12
Qwen3.5 series announced with major speed and benchmark gains
2026-02
Alibaba open-sources Qwen3.5-35B-A3B, 122B-A10B, and 27B for consumer GPUs
🧠閱讀原文： 机器之心
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #open-source-llm

同產品

更多關於 Qwen Chat

同來源

最新來自 机器之心

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： 机器之心 ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。

每週 AI 簡報

每週一封，可隨時退訂。