North Small Translate評測：翻譯贏Google卻卡授權 | North Small Translate Review: Beats Google, Locked License

9月 12, 2026

By Kit 小克 | AI Tool Observer | 2026-09-13

🇹🇼 North Small Translate評測：翻譯贏Google卻卡授權

Cohere 於 2026 年 9 月 10 日發表開權重翻譯模型 North Small Translate，218B 參數的 MoE 架構，官方數據在 WMT26 評測拿下 83.6 分，超越 DeepL NextGen 的 81.37 分與 Google Translate 的 68.20 分，是這週開發者圈討論度最高的 AI 話題之一。但看完授權條款後，故事沒那麼單純。

實測數據：贏在哪裡

North Small Translate 支援 50 種語言，官方公布的分數確實亮眼：

整體 WMT26 分數：83.6（多輪版本 84.36），對比 DeepL NextGen 81.37、Google Translate 68.20
長文件翻譯（整本書章節）：48.9 分，Google Translate 只有 21.3
非歐語系優勢明顯：南亞、中東北非地區贏 DeepL 約 8-10 分，東南亞贏 4-5 分

要注意的是，這些分數是 Cohere 自行公布、由 GPT-5.6-Sol 當裁判評分，不是第三方獨立驗證，實際品質建議自己餵幾段中文法律合約或口語對話測試看看。

架構：218B 沒有想像中肥

North Small Translate 是稀疏 MoE 架構，總參數 218B，但每次推論只啟用 25B（128 個專家取 8 個），實際運算量接近中型模型。官方測到的輸出速度是每秒 112 token，比 Gemma 4 31B 的 81 token 還快，換算下來單次翻譯任務成本只要 0.000676 美元。

想串接前先看授權

這是整個 North Small Translate 故事裡最容易被忽略的一段：

Hugging Face 上的權重是 CC BY-NC 4.0，只能研究、非商用，下載還要先留聯絡資訊給 Cohere
商用得走 Cohere Model Vault 另外簽約，不是下載就能用在產品裡
自架需求不小：BF16 版本要 4 張 B200 或 8 張 H100，量化到 4-bit 也要 1 張 B200 起跳，個人或小團隊基本架不動
免費管道是 Cohere Chat V2 API，額度內可以先玩玩看，不用自己扛硬體

誰該試試看

如果你的產品需要處理南亞、中東、東南亞語言的翻譯，或常常要處理長文件（合約、書籍、技術文件），North Small Translate 的分數差距在這些場景最明顯，值得評估。但如果只是想要「免費本地跑」，授權和硬體門檻會先讓你卻步——先用 API 試水溫是比較實際的路線。

好不好用，試了才知道。

🇺🇸 North Small Translate Review: Beats Google, Locked License

Cohere released its open-weight translation model North Small Translate on September 10, 2026 — a 218B-parameter MoE model that claims a WMT26 score of 83.6, beating DeepL NextGen's 81.37 and Google Translate's 68.20. It's one of the most-discussed AI releases this week among developers. But the license fine print changes the picture.

The Benchmark Numbers

North Small Translate covers 50 languages, and Cohere's published numbers are genuinely strong:

Overall WMT26 score: 83.6 (84.36 with the multi-pass agentic variant), vs. DeepL NextGen's 81.37 and Google Translate's 68.20
Long-document translation (book-chapter length): 48.9 vs. Google Translate's 21.3
Biggest edge in non-European languages: 8-10 points ahead of DeepL in South Asia and MENA, 4-5 points in Southeast Asia

Worth flagging: these are vendor-reported scores, judged by GPT-5.6-Sol rather than an independent third party. Run your own test set — a legal contract, a casual conversation — before trusting the leaderboard.

218B Isn't as Heavy as It Sounds

North Small Translate is a sparse MoE model: 218B total parameters, but only 25B active per token (8 of 128 experts). That keeps inference closer to a mid-sized dense model. Cohere reports 112 tokens/second throughput, faster than Gemma 4 31B's 81 tokens/second, at roughly /bin/zsh.000676 per translation task.

Check the License Before You Build on It

This is the part easiest to miss:

The Hugging Face weights are CC BY-NC 4.0 — research and non-commercial use only, and downloading requires sharing contact info with Cohere
Commercial use requires a separate agreement through Cohere's Model Vault — you can't just ship the downloaded weights in a product
Self-hosting isn't cheap: BF16 needs 4x B200 or 8x H100 GPUs; even 4-bit quantization needs at least 1x B200
The free path is Cohere's Chat V2 API, usable within rate limits without touching any hardware

Who Should Actually Try It

If your product handles South Asian, MENA, or Southeast Asian language pairs, or regularly translates long documents like contracts or technical manuals, North Small Translate's edge shows up exactly there and is worth evaluating. But if you're hoping to run it free and local, the license and hardware requirements will stop you fast — testing through the API first is the more realistic path.

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

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