Glonce

Tencent releases Hy4 preview, a 770B/49B open-source LLM with 1M+ tokens of context

modelsThe Glonce editorsHacker News🔥79Glonce rating79Impact75Novelty60Relevance100Surprise40Players70Credibility85
Tencent releases Hy4 preview, a 770B/49B open-source LLM with 1M+ tokens of context

Tencent has released and open-sourced Tencent Hy4 preview, a large language model with 770 billion total parameters and 49 billion active parameters, and a context window exceeding 1 million tokens. Tencent built it for real-world productivity work, citing strong performance on coding, office tasks and scientific research, and calls it part of the top tier of open-source models, though it does not name a specific benchmark or leaderboard behind that ranking. The model is open-sourced, and Tencent says it can also be reached through WorkBuddy, CodeBuddy, Yuanbao, ima and other Tencent products, or via API through Tencent Cloud TokenHub and OpenRouter.

On launch, Hy4 preview is free to use on WorkBuddy and CodeBuddy for two weeks; separately, Tencent has extended free access to its earlier Hy3 model on both platforms through September 30, with no year stated. API pricing for Hy4 preview is USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits.

Tencent says Hy4 preview was trained on data co-created with its own experts across software engineering, gaming, finance, security and other domains, in close coordination with WorkBuddy and CodeBuddy, and describes gains by domain. In software engineering: stronger understanding, planning, debugging and validation on long-context coding work, plus better front-end visual and interaction quality. In office and analytical work: better handling of complex working environments and financial analysis, and support for the full workflow from processing information to producing documents, spreadsheets and presentations. In game development: generating a playable prototype from a single natural-language request, then refining it through multi-turn interaction with game engines. In scientific research: stronger reasoning on complex problems, including AI research and development itself, molecular dynamics simulation, condensed-matter physics and fundamental mathematics.

Tencent also says Hy4 preview took part in its own development for the first time, contributing to automated optimization of training methods, data strategies, evaluation frameworks and low-level operators: the model proposed approaches, ran experiments and iterated on the results, with the resulting code, logs and feedback feeding into further rounds. Tencent calls this 'an early-stage recursive self-improvement loop.' Separately, the model analyzed bottlenecks in its own inference system and ran several rounds of optimization, including operator fusion and communication tuning, which Tencent says raised end-to-end throughput by 31.8% over baseline, consistently across different context lengths and concurrency levels.

In an internal blind evaluation that Tencent ran with 163 experts across 203 engineering tasks, Hy4 preview averaged 2.99 out of 4.00, ahead of GLM-5.3 (2.92) and Kimi K3 (2.94); Tencent's own wording calls the lead slight. The company frames the release as part of a preview-first approach under its Hunyuan effort, where real-world feedback folds back into research and development, and says the next batch of Hy4-series models is expected soon, without giving a date.

Key facts

  • Tencent released and open-sourced Hy4 preview: 770 billion total parameters, 49 billion active parameters, and a context window exceeding 1 million tokens.
  • Hy4 preview is free on WorkBuddy and CodeBuddy for two weeks after launch; free access to the earlier Hy3 model on both apps has been extended through September 30.
  • In Tencent's own internal blind evaluation (163 experts, 203 engineering tasks), Hy4 preview averaged 2.99 out of 4.00, ahead of GLM-5.3 (2.92) and Kimi K3 (2.94) by what Tencent calls a slight margin.
  • Tencent says the model helped automate parts of its own training pipeline, including data strategies, evaluation frameworks and low-level operators, for the first time, calling it an early-stage recursive self-improvement loop, and separately optimized its own inference system for a 31.8% throughput gain over baseline.
  • API pricing is USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits.

Why it matters

Hy4 preview is a large openly released model, 770 billion total parameters with a context window past 1 million tokens, and Tencent positions it against other open-source models, GLM-5.3 and Kimi K3, rather than closed frontier labs. The more unusual claim is that the model contributed to its own development: Tencent says Hy4 preview helped automate parts of training-method and data-strategy optimization, and separately analyzed and sped up its own inference system. Tencent calls this 'an early-stage recursive self-improvement loop,' a concrete but narrow example of a model assisting its own engineering process rather than a claim about broad autonomy.

Who it affects

Developers and companies building on open-source models, since Hy4 preview is open-sourced and also reachable through Tencent Cloud TokenHub and OpenRouter. Users of Tencent's own WorkBuddy and CodeBuddy get free access to the new model for two weeks, plus continued free access to Hy3 through September 30; Yuanbao, ima and other Tencent products can also reach it. The comparison in Tencent's internal evaluation also puts GLM-5.3 and Kimi K3, the two other open-source models it benchmarked against, in the spotlight.

How to use it

Hy4 preview is open-sourced and reachable through WorkBuddy, CodeBuddy, Yuanbao, ima and other Tencent products, or via API through Tencent Cloud TokenHub and OpenRouter. It is free on WorkBuddy and CodeBuddy for two weeks after launch; API pricing is USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits. Tencent has also extended free access to its earlier Hy3 model on WorkBuddy and CodeBuddy through September 30.

How solid is it

Every figure here comes from Tencent's own release; no independent researcher, executive or outside reviewer is named in the source, and no outside benchmark or leaderboard evaluated the model. The internal blind evaluation Tencent cites, 163 experts scoring 203 engineering tasks, was run and scored by Tencent itself, with no published rubric beyond those counts, and Tencent's own wording calls its lead over GLM-5.3 and Kimi K3 slight rather than decisive. Describing the model as ranking among the top tier of open-source models is likewise Tencent's own characterization, not tied to a named leaderboard. Tencent also says the model was expanded significantly in model size, context length and data volume compared with Hy3, but gives no Hy3 figures to measure that claim against.

Risks and caveats

The recursive self-improvement loop Tencent describes is narrow and self-reported: the model helped tune specific steps, data strategies, evaluation frameworks, low-level operators, within a process Tencent itself still ran and labeled early-stage, not autonomous general improvement. The 31.8% inference speedup and the 2.99 out of 4.00 evaluation score both come from Tencent's own internal testing, with no outside party checking either figure. No specific launch date is given for Hy4 preview, and no year is stated for the September 30 Hy3 deadline. The source does not explain why total parameters (770 billion) and active parameters (49 billion) differ, so the model's exact architecture is not established here.