Tencent has released and open-sourced Hy4 preview, a large language model featuring 770 billion total parameters and 49 billion active parameters alongside a context window exceeding one million tokens. The release uses a sparse mixture-of-experts design to process long continuous prompts while routing inference through a fraction of its total parameter count.

Practitioners can inspect the open-source release directly or test it through hosted consumer and developer channels without managing dedicated infrastructure:

  • Workplace and developer tools: Tencent announced access through WorkBuddy, CodeBuddy, Yuanbao, and ima. At launch, it offered two weeks of free access on WorkBuddy and CodeBuddy following launch, while extending free access to Hy3 on both tools through September 30, 2026.
  • Hosted API platforms: External developers can query the model via Tencent Cloud TokenHub and OpenRouter.
  • API pricing: Standard API rates are set at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-hit tokens.

Tencent framed the model around practical production workflows rather than general text generation. According to company claims, Hy4 preview handles code planning, debugging, and validation for long-context software engineering, while supporting full office workflows spanning documents, spreadsheets, and slide decks. Tencent also reported capabilities in fundamental mathematics, molecular dynamics simulation, condensed-matter physics, and game development, asserting that the model can generate a playable game prototype from a single natural-language request for further multi-turn engine refinement.

The vendor also claimed Hy4 preview contributed to its own development cycle by analyzing training routines and low-level inference bottlenecks. Tencent said the model's automated optimizations in operator fusion and communication raised end-to-end inference throughput by 31.8% over an unspecified internal baseline.

In a company-run blind evaluation across 203 engineering tasks scored by 163 internal experts, Tencent reported an average score of 2.99 out of 4.00 for Hy4 preview, compared to 2.92 for GLM-5.3 and 2.94 for Kimi K3. Tencent stated that additional models in the Hy4 series are expected to roll out soon, without setting specific rollout dates in the announcement.