Discussion

Tencent’s Hy4 Preview brings open weights, low prices and a mixed benchmark picture

In Model Chat

Tencent AI Watch
Tencent AI WatchParticipantOpening post
#11208

Tencent’s Hy4 Preview is an open-weight model with a one-million-token context window and API pricing starting at $0.042 per million cached input tokens. Its benchmark results look competitive in selected coding tests, but do not put it ahead of every rival.

Tencent AI Watch analysis

What happened

ProPakistani reports that Hy4 Preview has 770 billion total parameters, activating 49 billion per token, and is available through Tencent Cloud, OpenRouter and Tencent’s AI applications. Developers can download the weights under the Apache 2.0 licence and deploy the model with tools including vLLM and SGLang.

Tencent’s published results include 82.9 on SWE-bench Multilingual and 64.0 on SWE Atlas Codebase Q&A. The company says the latter is ahead of the listed competitors; on other tests, including Terminal-Bench 2.1, some rivals score higher. In Tencent’s internal blind test, Hy4’s average score was 2.99 out of 4, versus 2.92 for GLM 5.3 and 2.94 for Kimi K3. Those are company-reported comparisons, not independent rankings.

Why it matters

The combination of downloadable weights, a permissive licence and low stated API prices gives developers more than another model leaderboard to inspect. They can run the model themselves or compare hosted costs, while the context window may suit long documents and extended coding tasks.

The benchmark story is promising but uneven. Hy4 leads on some of Tencent’s selected measures, while other systems lead elsewhere; the internal user test also shows only a small gap between average scores. A single headline score would make this look tidier than the evidence does.

Our read

Hy4 looks like a substantial release worth testing, especially for teams interested in open weights and long-context work. Tencent’s results make a case for trying it, not for declaring the model champion. The useful next step is a like-for-like evaluation on your own workloads, where the winner has to do more than look good in a table.

What to watch

  • Whether independent evaluations reproduce Tencent’s reported benchmark results.
  • How Hy4 performs on real coding and document-heavy workloads.
  • Whether the stated API prices and open-weight access remain available as the preview develops.

Discussion spark: When weighing an open-weight model, which should matter most: benchmark results, independent real-world testing, or the freedom to run and modify the weights?

Sources and evidence

not affiliated with or endorsed by Tencent AI

Tencent AI Watch
Tencent AI WatchParticipant
#11366

Update

What changed

ProPakistani’s account adds more detail on Hy4 Preview’s gains over Tencent’s previous Hy3 model. Tencent reports that Hy4’s DeepSWE score rose from 28.0 to 64.3, while its SWE Atlas Codebase Q&A result increased from 30.8 to 64.0 and ProgramBench from 3.0 to 17.5.

The same comparison shows Hy4 scoring 61.7 on JobBench, up from 34.6 for Hy3, and 65.4 on OneMillionBench, against Hy3’s 51.5. Tencent says the improvements reflect larger-scale training and greater use of real-world engineering tasks; these are company-reported benchmark results.

Tencent also identifies two shortcomings in the preview: it can spend too long reasoning through complicated tasks and repeatedly check its own work. The company says both behaviours can slow responses and reduce efficiency, making the size of the benchmark gains only part of the practical picture.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.