Tencent released Hy4 Preview on 28 August 2026 — an open-weights MoE model with 770B total parameters and 49B active. It more than doubles Hy3 in scale and is the first Tencent model that helped optimise the infrastructure it runs on.
Key takeaways
- MoE with 770B total and 49B active parameters — Hy3 had 295B and 21B
- Context window above 1M tokens, four times the 256K in Hy3
- Weights on Hugging Face: 1.56 TB, Apache 2.0 license, no vision modality
- 2.99/4.00 across 203 internal engineering tasks, ahead of Kimi K3 (2.94) and GLM-5.3 (2.92)
- API via Tencent Cloud TokenHub and OpenRouter, from $0.834 per million input tokens
Scale grows faster than the cost per request
Of the 770B parameters in the full model, each token engages only 49B. The active share fell from roughly 7 percent in Hy3 to under 6.4 percent, buying capacity without a proportional rise in inference cost. The cost lands elsewhere: the weights take 1.56 TB against 598 GB for Hy3, ruling out running it outside a server cluster. It is text-only, and the chat template offers two reasoning modes?high / no_think reasoning modes: Two settings of the reasoning_effort parameter in the Hy4 chat template. The default high runs an extended chain of thought, while no_think turns that step off and returns a direct answer.: the default “high” and “no_think”.
A lead measured in hundredths of a point
An internal evaluation had 163 experts run 203 engineering tasks, and Hy4 Preview reached 2.99 out of 4.00 — just ahead of Kimi K3 (2.94) and GLM-5.3 (2.92). That gap comes from a vendor-designed test, and independent results are not yet available. Tencent’s other claim is more interesting: Hy4 improved the inference infrastructure itself, lifting throughput by 31.8 percent.
| Model | Total params | Active params | Context window | Score (203 tasks) |
|---|---|---|---|---|
| Hy4 Preview | 770B | 49B | over 1M tokens | 2.99 |
| Hy3 | 295B | 21B | 256K tokens | — |
| Kimi K3 | — | — | — | 2.94 |
| GLM-5.3 | — | — | — | 2.92 |
Availability and pricing
The model runs inside WorkBuddy, CodeBuddy, Yuanbao and ima, free on the first two for two weeks from launch. The API runs through Tencent Cloud TokenHub and OpenRouter, with a separate, far lower rate for cache-hit?cache-hit: A prompt fragment the provider has already processed and keeps in its cache. Those tokens are billed at a far lower rate than fresh input tokens. tokens.
Hy4 Preview API pricing per million tokens:
Why it matters
Open weights under Apache 2.0 at this scale shift the balance: teams that until now chose between a closed API and a much weaker local model get a third option. Yet 1.56 TB is a barrier that makes this openness practical mainly for infrastructure operators, not individual companies. The claimed self-optimisation needs independent confirmation before it becomes a market argument.
What next?
- The “preview” label marks a non-final release — no date given for a stable Hy4
- Free access in WorkBuddy and CodeBuddy expires around 11 September 2026
- Weights are public and the model is on OpenRouter, so the 2.99/4.00 score can be independently checked
Sources
- Tencent — Tencent Releases and Open-Sources Tencent Hy4 Preview
- Simon Willison's Weblog — Hy4
- Hugging Face — tencent/Hy4-Preview
