Tencent Open-Sources Hy4 Preview: 770B, 49B Active
The open-weight model runs 1M tokens of context, weighs 1.56 terabytes on Hugging Face and lists input tokens at $0.834 per million.
Illustration: a night-time data center hall with endless rows of servers and thin blue status lights.
Tencent has released Hy4 Preview as an open-weight, text-only model with 770 billion total parameters, 49 billion active per token and a 1-million-token context window.
At a glance
- 770 billion total parameters, 49 billion active per token; mixture-of-experts, text input only, no vision.
- Context window: 1 million tokens. The Hugging Face download totals 1.56 terabytes.
- Predecessor Hy3, July 2026: 295 billion parameters, 21 billion active, 256,000-token context, 598 gigabytes.
- Tencent's list price: $0.834 per million input tokens, $2.501 output, $0.042 on cache hits.
- Internal review: 2.99 of 4.00 from 163 experts across 203 tasks; GLM-5.3 scored 2.92, Kimi K3 2.94.
Tencent has published Hy4 Preview, an open-weight model with 770 billion total parameters that activates 49 billion of them per token. It accepts text only — there is no image input. The context window is 1 million tokens, and the Hugging Face download comes to 1.56 terabytes.
How it compares with Hy3
Hy3 arrived in July 2026 at 295 billion total parameters, 21 billion active, a 256,000-token context window and 598 gigabytes on disk. Roughly six weeks later, Hy4 Preview more than doubles the total parameter count and stretches the context window close to fourfold.
The numbers Tencent publishes
Tencent had 163 experts score 203 engineering tasks drawn from its own teams' work. Hy4 Preview averaged 2.99 out of 4.00, ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. That is a vendor-run evaluation rather than an independent benchmark, and a gap of 0.05 points is thin enough to warrant caution.
Tencent also reports a 31.8 percent gain in inference throughput, which it credits to the model optimizing its own serving path. The announcement does not describe how that figure was measured.
Pricing and where to run it
Through Tencent Cloud TokenHub and OpenRouter, the listed rates are $0.834 per million input tokens and $2.501 per million output tokens, dropping to $0.042 per million on cache hits. WorkBuddy and CodeBuddy were free for the first two weeks after launch, and the Yuanbao and ima apps carry the model as well.
What the sources do not settle
Product Hunt lists the model under Apache 2.0. Tencent's own announcement says open source without naming a license, so across the three sources read for this article the license designation is not firmly established.
The release date differs too: Tencent dates it to August 28, 2026, while the other two sources give August 29, 2026. On behavior, the team acknowledges on Product Hunt that the preview tends to over-verify its own work, and the model exposes two reasoning settings — "high" by default and "no_think" with reasoning switched off.
FAQ
How many parameters does Tencent Hy4 Preview have?
770 billion in total, with 49 billion active per token under its mixture-of-experts design.
Does Hy4 Preview support image input?
No. It is a text-only model; Tencent ships no vision capability in this preview.
How much does the Hy4 Preview API cost?
Tencent lists $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million on cache hits.
