Tencent Hy4評測：770B開源模型贏過GLM-5.3與Kimi K3 | Tencent Hy4 Review: 770B Open Model Beats GLM-5.3

By Kit 小克 | AI Tool Observer | 2026-08-31

🇹🇼 Tencent Hy4評測：770B開源模型贏過GLM-5.3與Kimi K3

騰訊(Tencent)在8月28日無預警開源了旗艦模型 Tencent Hy4 Preview，770B參數的MoE（混合專家）架構，直接用Apache 2.0授權釋出權重——這代表任何人都能免費下載、商用、魔改，不用簽合約也不用付授權金。更狠的是，騰訊自己公布的內部盲測數字顯示，Hy4在多項任務上贏過現在最紅的GLM-5.3和Kimi K3，SWE-bench多語言測試甚至超車DeepSeek V4 Pro。這篇文章帶你看規格、效能數字跟怎麼用。

什麼是 Tencent Hy4 Preview？

Tencent Hy4 Preview 是騰訊釋出的旗艦級開源大型語言模型，採用MoE架構，總參數770B、實際運算時只啟用49B，搭配超過1M token的超長上下文，主打程式開發、辦公文件生成、遊戲原型與科學研究等生產力場景。

架構：78層，第一層為標準dense FFN，其餘77層改為MoE，每層256個路由專家+1個共享專家，每個token啟動top-8路由專家
授權：Apache 2.0，允許商用、修改、再散布
釋出管道：Hugging Face、ModelScope、GitCode、CNB，共131個權重檔案，含微調腳本

Hy4 效能真的贏過 GLM-5.3 和 Kimi K3 嗎？

根據騰訊公布的內部盲測，163位工程師針對203項真實任務評分，Hy4平均拿下2.99分（滿分4.00），些微領先GLM-5.3（2.92）和Kimi K3（2.94），在46.8%案例贏過GLM-5.3、51.2%案例贏過Kimi K3。

SWE-bench Multilingual：Hy4 82.9分，GLM-5.3 81.3分，Kimi K3 80.8分，DeepSeek V4 Pro 77.3分
輸出成本比Kimi K3便宜約82%

要注意的是，這些是騰訊自己做的內部評測，還沒有大量第三方獨立跑分驗證，數字看看就好，實際好不好用還是得自己上手測。

Hy4 怎麼用？多少錢？

Hy4可以直接下載權重自架，或透過OpenRouter、騰訊雲TokenHub呼叫API，價格是每百萬input token約0.83美元、output token約2.5美元，快取命中每百萬token只要0.042美元，算是同等級模型裡便宜的一檔。

小克怎麼看

Hy4最大的意義不是「又一個模型贏過誰」，而是770B這個量級加上Apache 2.0全開放授權，把開源模型的天花板又往上推了一層。中國實驗室這幾個月一個接一個開源旗艦模型（DeepSeek、GLM、Kimi、現在騰訊），對比美國幾家閉源當道，這個趨勢值得繼續盯。但騰訊的benchmark是自己發布的，實測落差多大、社群跑分怎麼說，還得再等幾週。好不好用，試了才知道。

🇺🇸 Tencent Hy4 Review: 770B Open Model Beats GLM-5.3

Tencent quietly open-sourced its flagship Tencent Hy4 Preview model on August 28, a 770B-parameter Mixture-of-Experts (MoE) system released under a full Apache 2.0 license — free to download, modify, and use commercially, no contract or royalty required. Tencent's own internal blind tests claim Hy4 beats the currently red-hot GLM-5.3 and Kimi K3 on several tasks, and even edges out DeepSeek V4 Pro on SWE-bench Multilingual. Here's what the specs, benchmarks, and pricing actually look like.

What Is Tencent Hy4 Preview?

Tencent Hy4 Preview is Tencent's flagship open-weight large language model: an MoE architecture with 770B total parameters but only 49B active per token, paired with a context window over 1M tokens, aimed at coding, office document generation, game prototyping, and scientific research.

Architecture: 78 layers — the first is a standard dense FFN, the remaining 77 use MoE with 256 routed experts plus 1 shared expert; each token activates the top-8 routed experts
License: Apache 2.0 — commercial use, modification, and redistribution allowed
Availability: Hugging Face, ModelScope, GitCode, and CNB, with 131 weight files plus fine-tuning scripts

Does Hy4 Actually Beat GLM-5.3 and Kimi K3?

In Tencent's internal blind evaluation, 163 engineers scored 203 real-world tasks; Hy4 averaged 2.99 out of 4.00, slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94), winning 46.8% of head-to-head cases against GLM-5.3 and 51.2% against Kimi K3.

SWE-bench Multilingual: Hy4 scores 82.9, versus GLM-5.3 (81.3), Kimi K3 (80.8), and DeepSeek V4 Pro (77.3)
Output cost runs roughly 82% cheaper than Kimi K3

Worth flagging: these are Tencent's own internal numbers, not yet backed by extensive independent benchmarking — treat them as a starting point, not gospel.

How Much Does Hy4 Cost and Where Can You Run It?

You can self-host the downloaded weights, or call it via OpenRouter or Tencent Cloud TokenHub at roughly $0.83 per million input tokens and $2.50 per million output tokens, with cache-hit tokens at just $0.042 per million — cheap for a model at this scale.

Kit's Take

The real story here isn't "another model beats another model" — it's that a 770B-parameter flagship dropping under full Apache 2.0 keeps pushing the open-source ceiling higher. Chinese labs have open-sourced flagship after flagship this year (DeepSeek, GLM, Kimi, now Tencent) while several US labs stay closed — that gap is worth watching. But since the benchmarks are self-reported, how it holds up against independent testing and community usage is still an open question. 好不好用，試了才知道 — you only know if it's good once you've actually tried it.

Sources / 資料來源

常見問題 FAQ

Tencent Hy4 Preview 是免費的嗎？

權重完全免費下載、可商用，但透過OpenRouter或騰訊雲TokenHub的API呼叫仍要按token計費，約每百萬input token 0.83美元。

Hy4的效能數字可信嗎？

目前的benchmark都是騰訊自行公布的內部盲測與SWE-bench成績，還沒有足夠的第三方獨立驗證，建議實測後再下結論。

Hy4跟GLM-5.3、DeepSeek有什麼差別？

三者都是中國實驗室開源的旗艦MoE模型，Hy4參數規模最大（770B）、上下文最長（1M token），定價也具競爭力，但各家benchmark互有領先。

在哪裡可以下載Hy4？

Hugging Face、ModelScope、GitCode、CNB都有釋出權重與微調腳本。

Hy4的授權可以商用嗎？

可以，採用Apache 2.0授權，允許商用、修改與再散布，不需另外簽約。

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