跳到主要內容

Qwen3.8-Max評測:阿里2.4兆參數模型開源迎戰Fable 5 | Qwen3.8-Max Review: Alibaba's 2.4T Model Rivals Fable 5

By Kit 小克 | AI Tool Observer | 2026-08-07

🇹🇼 Qwen3.8-Max評測:阿里2.4兆參數模型開源迎戰Fable 5

Qwen3.8-Max是阿里巴巴8月3日發布的最新旗艦模型,總參數量來到2.4兆、每次推理啟用950億參數的MoE架構,官方公布的評測數字直接對標Claude Fable 5與GPT-5.6。這是阿里近期少數願意開源的Max等級模型,也讓「中國模型追平美系旗艦」的話題再度成為焦點。

Qwen3.8-Max規格與定價多少?

Qwen3.8-Max是2.4兆參數的混合專家(MoE)模型,實際啟用參數約950億,支援文字、圖片、影片輸入,單次可處理高達100萬token的上下文。API目前已上線阿里雲Model Studio,價格為每百萬輸入token 2美元、輸出token 6美元,快取token則是每百萬0.25美元,同時提供OpenAI相容與Anthropic相容介面,理論上把Claude Code或Codex的設定改個幾行就能接上。

Qwen3.8-Max真的贏過Fable 5和GPT-5.6嗎?

阿里公布的內部評測顯示,Qwen3.8-Max在OSWorld-Verified、Terminal-Bench 2.1等代理型任務上分數領先GPT-5.6與Claude Fable 5,在Arena.AI的文字類排行榜上也一舉衝上中國模型第一名。但要注意這些數字目前全部來自官方自報,第三方尚未有機會用開源權重獨立覆現,加上不同媒體整理出來的分數版本差異不小,建議先當參考、別直接當成定論。

開源時程與對開發者的實際意義

阿里表示Qwen3.8-Max的開源權重會「下週」釋出,這將是Qwen系列第一個Max等級的開源模型,也代表阿里從先前把旗艦模型收緊為閉源,重新轉向開源策略。對開發者來說重點有三:

  • 先API後開源:現在只能透過付費API測試,真正能自架、微調要等權重釋出,屆時才有機會驗證官方數字是否誇大。
  • 定價具競爭力:每百萬token 2/6美元的組合,比多數美系旗艦模型便宜,適合先拿來做大量呼叫的原型測試。
  • 相容層省事:OpenAI與Anthropic雙相容介面,讓既有的agent workflow遷移成本降低,但複雜推理、安全性仍建議實測再上生產環境。

常見問題 FAQ

Q: Qwen3.8-Max什麼時候開源?
A: 阿里表示開源權重將於發布後「下週」釋出,是Qwen系列首個開源的Max等級模型。

Q: Qwen3.8-Max的API價格是多少?
A: 每百萬輸入token 2美元、輸出token 6美元,快取token每百萬0.25美元。

Q: Qwen3.8-Max真的比Fable 5強嗎?
A: 目前只有阿里官方自報的評測數字,尚未有開源權重可供第三方獨立驗證,建議保守看待。

Q: 一般開發者現在能用嗎?
A: 可以透過阿里雲Model Studio的API直接呼叫,支援OpenAI與Anthropic相容介面,遷移成本不高。

好不好用,試了才知道。


🇺🇸 Qwen3.8-Max Review: Alibaba's 2.4T Model Rivals Fable 5

Qwen3.8-Max, Alibaba's newest flagship model released August 3, is a 2.4-trillion-parameter mixture-of-experts model that Alibaba claims matches or beats Claude Fable 5 and GPT-5.6 on several benchmarks. It's also one of the few Max-tier Qwen models Alibaba plans to open-source, reviving the "can a Chinese model really catch up" conversation.

Qwen3.8-Max specs and pricing

Qwen3.8-Max is a 2.4T-parameter MoE model activating roughly 95B parameters per token, handling text, image, and video input with up to 1 million tokens of context. It's live now on Alibaba Cloud's Model Studio at $2 per million input tokens and $6 per million output tokens, with cached tokens at $0.25/M. Both OpenAI-compatible and Anthropic-compatible endpoints are available, so pointing Claude Code or Codex at it is mostly a config change.

Does it actually beat Fable 5 and GPT-5.6?

Alibaba's own benchmarks show Qwen3.8-Max ahead of GPT-5.6 and Claude Fable 5 on agentic evals like OSWorld-Verified and Terminal-Bench 2.1, and it jumped to the top spot among Chinese models on Arena.AI's text leaderboard. The catch: every number so far is self-reported. Nobody outside Alibaba has been able to independently reproduce these scores with open weights yet, and different outlets are already reporting inconsistent figures — treat the headline numbers as a starting point, not a verdict.

Open-source timeline and what it means for developers

Alibaba says open weights will follow next week, making Qwen3.8-Max the first Max-class Qwen model to go open-source — a reversal after Alibaba kept its most recent flagships closed. Three practical takeaways:

  • API first, weights later: You can only test via the paid API right now. Self-hosting or fine-tuning — and truly validating the benchmark claims — has to wait for the weight release.
  • Competitive pricing: $2/$6 per million tokens undercuts most US flagship models, making it worth a shot for high-volume prototyping.
  • Compatible endpoints cut migration cost: Dual OpenAI/Anthropic compatibility lowers the switching cost for existing agent workflows, but test complex reasoning and safety behavior before touching production.

FAQ

Q: When does Qwen3.8-Max go open source?
A: Alibaba says weights will ship next week after the August 3 launch — the first Max-tier Qwen model to be open-sourced.

Q: How much does the Qwen3.8-Max API cost?
A: $2 per million input tokens, $6 per million output tokens, $0.25 per million cached tokens.

Q: Does Qwen3.8-Max really outperform Fable 5?
A: Only on Alibaba's own self-reported benchmarks so far — no independent verification exists until open weights ship.

Q: Can developers use it today?
A: Yes, via Alibaba Cloud's Model Studio API, with OpenAI- and Anthropic-compatible endpoints for low migration cost.

好不好用,試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code