Qwen3.8-Max評測：阿里2.4兆參數開源模型，一般人跑不動 | Qwen3.8-Max Review: Alibaba's 2.4T Model No One Can Run

By Kit 小克 | AI Tool Observer | 2026-08-26

🇹🇼 Qwen3.8-Max評測：阿里2.4兆參數開源模型，一般人跑不動

Qwen3.8-Max 評測來了——阿里巴巴這次直接端出2.4兆參數的開源模型，一口氣躋身史上最大開源權重模型之一，比 DeepSeek V4-Pro（約1.6兆）大上不少，只輸給 Kimi K3 的2.8兆。但比起參數量，更值得關注的是：這顆模型一般人根本跑不動。

Qwen3.8-Max 是什麼？

Qwen3.8-Max（內部代號 Qwen3.8-2.4T-A95B）是阿里巴巴 Qwen 團隊發表的混合專家（MoE）模型，總參數 2.4 兆，但每次推理只啟用約 950 億參數。它原生支援多模態，能吃文字、圖片、影片輸入，輸出文字，把視覺理解直接放進執行迴圈裡，而不只是單向辨識圖片。

規格上有什麼亮點？

上下文視窗：最長支援 100 萬 token
輸出長度：最長 12.8 萬 token
架構：細粒度 MoE，混合全注意力與線性注意力
主打用途：複雜推理與代理（agentic）任務

Qwen3.8-Max 能在本機跑嗎？

理論上開源，實際上跑不動。官方 BF16 權重總大小約 4.9TB，FP8 版本約 2.5TB，就算壓到 4-bit 量化也要 1.3～1.45TB。硬塞進最省的量化版本仍需要 397GB 以上顯存，而且因為是 MoE 架構，即使每次只啟動一小部分參數，全部權重仍得留在高速記憶體裡待命——等於你得準備一台多卡伺服器，規格比你現在讀這篇文章用的電腦貴上好幾倍。

一般開發者該怎麼用 Qwen3.8-Max？

務實的做法是透過 QwenCloud 用 API 呼叫，而不是自己架伺服器。阿里目前先讓模型在雲端可用，開源權重預計稍後才釋出，但即使權重公開，能真正自架的也只有握有多卡叢集、且有明確理由要自己掌控推理堆疊的團隊或企業。對多數個人開發者跟中小團隊來說，Qwen3.8-Max 的「開源」比較像是「透明可審查」，不是「可以帶回家跑」。

常見問題（FAQ）

Qwen3.8-Max 比 DeepSeek V4-Pro 大嗎？

是的，Qwen3.8-Max 有 2.4 兆參數，大於 DeepSeek V4-Pro 的約 1.6 兆，但小於 Kimi K3 的 2.8 兆。

一般家用電腦能跑 Qwen3.8-Max 嗎？

不行。就算是壓縮過的 4-bit 量化版本也需要超過 397GB 顯存與多卡伺服器，遠超一般消費級硬體。

沒有多卡伺服器的開發者該怎麼用？

建議透過阿里雲的 QwenCloud API 直接呼叫，不需要自己準備硬體。

好不好用，試了才知道。

🇺🇸 Qwen3.8-Max Review: Alibaba's 2.4T Model No One Can Run

Qwen3.8-Max just landed, and Alibaba isn't messing around: this is a 2.4 trillion parameter open-weight model, instantly one of the largest open-weight releases ever — bigger than DeepSeek V4-Pro (~1.6T), just behind Kimi K3's 2.8T. But the parameter count isn't the real story. The real story is that almost nobody can actually run it.

What Is Qwen3.8-Max?

Qwen3.8-Max (internally Qwen3.8-2.4T-A95B) is Alibaba's new mixture-of-experts (MoE) model with 2.4 trillion total parameters, of which only about 95 billion activate per inference pass. It's natively multimodal — text, image, and video in, text out — with vision treated as part of the execution loop rather than a bolt-on.

What's Actually New in the Specs?

Context window: up to 1 million tokens
Output length: up to 128K tokens
Architecture: fine-grained MoE with hybrid full and linear attention
Focus: complex reasoning and agentic workloads

Can You Actually Run Qwen3.8-Max Locally?

Open-weight on paper, unrunnable in practice for almost everyone. The official BF16 weights total roughly 4.9TB, FP8 comes in around 2.5TB, and even 4-bit quantization still needs 1.3-1.45TB. Squeeze it into the leanest available quant build and you still need 397GB+ of fast memory across a multi-GPU server — because MoE architecture means all the weights have to sit ready in memory even though only a slice activates per token. You're looking at a server that costs more than the machine you're reading this on.

So How Should Developers Actually Use Qwen3.8-Max?

The practical move is calling it through Alibaba's QwenCloud API instead of self-hosting. Alibaba made the model cloud-accessible first, with open weights planned to follow — but even once weights are public, real self-hosting is only realistic for teams that already run multi-GPU clusters with a concrete reason to own the inference stack. For most individual developers and small teams, "open-weight" here means "auditable," not "something you bring home and run."

FAQ

Is Qwen3.8-Max bigger than DeepSeek V4-Pro?

Yes. Qwen3.8-Max has 2.4 trillion parameters versus DeepSeek V4-Pro's roughly 1.6 trillion, though it's smaller than Kimi K3's 2.8 trillion.

Can I run Qwen3.8-Max on a home PC?

No. Even the compressed 4-bit quantized build needs over 397GB of memory across a multi-GPU server, far beyond consumer hardware.

How should developers without a GPU cluster use it?

Call it through Alibaba's QwenCloud API directly — no hardware investment required.

好不好用，試了才知道 — good or not, you won't know until you try it.

Sources / 資料來源

常見問題 FAQ

Qwen3.8-Max 比 DeepSeek V4-Pro 大嗎？

是的，Qwen3.8-Max 有 2.4 兆參數，大於 DeepSeek V4-Pro 的約 1.6 兆，但小於 Kimi K3 的 2.8 兆。

一般家用電腦能跑 Qwen3.8-Max 嗎？

不行。就算是壓縮過的 4-bit 量化版本也需要超過 397GB 顯存與多卡伺服器，遠超一般消費級硬體。

Qwen3.8-Max 支援多長的上下文？

最長支援 100 萬 token 輸入，輸出上限為 12.8 萬 token。

沒有多卡伺服器的開發者該怎麼用 Qwen3.8-Max？

建議透過阿里雲的 QwenCloud API 直接呼叫，不需要自己準備硬體。

Qwen3.8-Max 是完全開源嗎？

目前先在雲端提供存取，阿里表示開源權重稍後才會釋出，屬於「先雲端、後開源」的做法。

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

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

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士