Tencent open-sources 770B-parameter Hy4 preview with over 1M token context

The release

Tencent has released and open-sourced Hy4 preview, a large language model it places among the strongest openly available models today. According to the company's announcement, which reached the Hacker News front page, the model has 770 billion total parameters, of which 49 billion are active at any given time, and it handles a context window of more than one million tokens. The target is not casual chat but practical output: software development, office work and scientific research are the headline use cases.

The preview label is intentional. Tencent says it ships models in preview first, gathers feedback from real-world usage, and folds that into official releases. Additional models in the Hy4 series are expected soon.

How Tencent rates it

To position the model against competitors, Tencent ran an internal blind evaluation involving 163 experts and 203 engineering tasks. Hy4 preview scored an average of 2.99 out of 4.00, edging out GLM-5.3 at 2.92 and Kimi K3 at 2.94, according to the company. Two caveats apply: the study was designed and run by Tencent itself, and the announcement does not describe an external audit, so the ranking should be read as a vendor claim rather than a settled result.

Task-level claims

The announcement breaks the model's improvements into several domains:

• Software engineering: better understanding, planning, debugging and validation on long-context development work, plus higher visual quality in front-end output.
• Office productivity: stronger analysis of complex working environments, financial analysis, data analysis and cross-document collaboration, covering the workflow from processing information through to producing documents, spreadsheets and slides.
• Game development: the model can generate a playable prototype from one natural-language prompt and cooperate with game engines, with further refinement through multi-turn conversation.
• Research: reported gains in AI research, molecular dynamics simulation, condensed-matter physics and fundamental mathematics.

Tencent attributes part of this to its training data, which was built with the help of in-house specialists in software engineering, gaming, finance and security, and to joint development with its own products, CodeBuddy and WorkBuddy.

The model helped build itself

The most unusual claim in the announcement concerns the model's own development. For the first time, Hy4 preview took part in automating parts of its training pipeline, including training methods, data strategy, evaluation frameworks and low-level operators. Tencent says the model proposed approaches, ran experiments and iterated on the results, with the resulting code, logs and feedback feeding into later rounds of exploration, forming what the company characterizes as an early loop of recursive self-improvement.

The model also examined bottlenecks in its own inference stack and applied several rounds of optimization, including operator fusion and communication tuning. Tencent reports a 31.8% increase in end-to-end throughput over the baseline, with gains holding across different context lengths and concurrency levels. As with the benchmark figures, these numbers come from the vendor without independent verification.

Availability and pricing

Hy4 preview is available as an open-source release and through several Tencent products: WorkBuddy, CodeBuddy, Yuanbao and ima. API access runs through Tencent Cloud TokenHub and OpenRouter. Access on WorkBuddy and CodeBuddy is free for two weeks following launch, and free access to the previous Hy3 model on both platforms has been extended to September 30.

On pricing, the API costs USD 0.834 per million input tokens, USD 2.501 per million output tokens and USD 0.042 per million tokens on cache hits, which Tencent frames as a way of keeping advanced AI affordable.

Why it matters

Hy4 preview is another data point in a clear trend: frontier-scale open models increasingly come from Chinese labs, and they arrive with capabilities and pricing that compete directly with closed alternatives. The 770B-total, 49B-active design combined with a one-million-token context window is squarely aimed at long, agentic work sessions, such as reading large codebases or analyzing many documents at once, rather than short chat exchanges.

The self-improvement claims also deserve attention. If a model genuinely contributes to its own training pipeline and speeds up its own inference infrastructure, even partially, the effect compounds over time: better models help build better models. Independent verification is still needed, and the preview status means the final release may differ from what ships now. Even so, the combination of open weights, a very large context window and low API pricing gives developers a serious new option to evaluate.