Qwen3.8-27B評測：Cerebras跑出1500 tok/s實測 | Qwen3.8-27B Review: 1,500 Tokens/Sec on Cerebras

9月 05, 2026

By Kit 小克 | AI Tool Observer | 2026-09-06

🇹🇼 Qwen3.8-27B評測：Cerebras跑出1500 tok/s實測

阿里巴巴Qwen團隊發布的Qwen3.8-27B，這款270億參數的多模態模型，本週在Hacker News衝上熱門榜首的原因不是模型本身的跑分，而是它被Cerebras晶圓級推理晶片扛起來之後，實測跑出每秒1500個token的輸出速度——差不多是主流GPU雲端服務的3到5倍。對需要即時多輪呼叫工具的AI Agent應用來說，這個數字比一堆基準測試分數都更有感。

Qwen3.8-27B到底是什麼模型

Qwen3.8-27B是一款27B參數的稠密（dense，非MoE）多模態模型，支援文字加PNG/JPEG圖片輸入，內建可調式推理（reasoning）模式、工具呼叫（tool calling）、串流輸出與結構化輸出，主打場景是agentic coding、工具鏈調用與長時間執行的研究型工作流程。相較於一味堆參數，這一版把重點放在「好伺服、好落地」——27B的體積剛好能被Cerebras的硬體吃到滿速，自架或走API的成本也不會太誇張。

Cerebras的晶圓級推理，1500 tok/s怎麼跑出來的

Cerebras的殺手鐧是晶圓級引擎（Wafer-Scale Engine），把數十萬個運算核心整合在單片矽晶圓上，省掉一般多GPU叢集在晶片之間搬資料造成的延遲瓶頸。這也是為什麼同一份Qwen3.8-27B權重，放到Cerebras上跑，輸出速度可以衝到1500 tokens/秒，比典型GPU雲端服務快上一大截。

定價與速率限制

免費試用層：64K context、32K最大輸出、每分鐘5次請求、每天100萬token上限
付費開發者層：128K context、40K最大輸出、每分鐘300次請求、無每日上限
定價：每百萬input token 0.99美元，output token 1.49美元

值不值得換

如果你的應用場景是即時對話、agent多輪呼叫工具、或需要在使用者等待時間內跑完長推理鏈，1500 tok/s的差距是肉眼可見的——原本要等5秒的回應，現在可能1秒內就跑完。但要注意：Cerebras目前是「特定模型上架」模式，不是你能隨便丟自己的模型上去跑，能不能吃到這速度，取決於Qwen3.8-27B剛好符合你的需求；如果你要的是更大模型（70B以上）或特定微調版，目前還沾不到這個加速。免費層每分鐘5次請求的限制，也讓它現階段比較適合拿來測試，還不是生產環境的量能。

好不好用，試了才知道。

🇺🇸 Qwen3.8-27B Review: 1,500 Tokens/Sec on Cerebras

Alibaba's Qwen team released Qwen3.8-27B, a 27-billion-parameter multimodal model, but the reason it hit the top of Hacker News this week isn't the benchmark card — it's that Cerebras put it into production at roughly 1,500 tokens per second, several times faster than typical GPU-cloud serving. For agentic apps chaining multiple tool calls in real time, that number matters more than another leaderboard score.

What Qwen3.8-27B Actually Is

Qwen3.8-27B is a dense (not MoE) 27B-parameter multimodal model that accepts text plus PNG/JPEG images, with configurable reasoning, tool calling, streaming and structured outputs. It's aimed at agentic coding, tool-chaining, and long-running research workflows. Rather than chasing raw parameter count, this generation looks tuned for practicality — 27B is small enough for Cerebras's chips to run at full speed, and cheap enough to serve without a runaway bill.

How Cerebras Hits 1,500 Tokens/Sec

Cerebras's edge is its Wafer-Scale Engine (WSE) — hundreds of thousands of compute cores etched onto a single silicon wafer, which removes the inter-chip communication bottleneck that slows down typical multi-GPU clusters. That architecture is why the same Qwen3.8-27B weights, served on Cerebras, can push out roughly 1,500 tokens per second, a multiple of what you'd normally see from GPU cloud endpoints.

Pricing and Rate Limits

Free trial: 64K context, 32K max output, 5 requests/min, 1M tokens/day cap
Paid developer tier: 128K context, 40K max output, 300 requests/min, no daily cap
Pricing: $0.99 per million input tokens, $1.49 per million output tokens

Is It Worth Switching

If your use case is real-time chat, an agent chaining multiple tool calls, or a long reasoning trace users are waiting on, 1,500 tok/s is a difference you can actually feel — a 5-second wait can drop to under a second. The catch: Cerebras runs a curated model catalog, not a bring-your-own-model service, so you only get this speed if Qwen3.8-27B happens to fit your needs. Need something bigger (70B+) or a specific fine-tune? It is not accelerated here yet. The free tier'''s 5-requests-per-minute cap also makes it better suited to testing than production traffic today.

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

搜尋此網誌

AI小貼士