Tencent's Largest Open-Weight Release to Date
On 28 August 2026, Tencent released and open-sourced Hy4 preview, the newest model in the Hunyuan family. Hy4 is a Mixture-of-Experts language model with 770 billion total parameters, of which only 49 billion are active per token — meaning the model routes each input through a subset of its parameter space rather than the full weight matrix. The release is licensed under Apache 2.0, making it freely available for commercial and research use without royalty obligations.
Architecture and What the Numbers Mean
A 770B MoE model activates roughly 6.4% of its parameters per forward pass. That ratio is the efficiency story: the model carries the representational depth of a 770 billion parameter system while incurring inference cost closer to a 49 billion parameter dense model. For teams running the model at scale, this matters because it directly affects the number of requests per second a given hardware cluster can serve. An FP8 quantised variant ships alongside the full-precision release, further reducing memory requirements for teams deploying on standard GPU infrastructure. The context window is 1 million tokens, placing Hy4 preview among the most context-capable open-weight models available as of late August 2026.
The Self-Optimisation Claim
The most technically interesting aspect of the Hy4 release is what Tencent says about how the model was built. The company reports that Hy4 preview autonomously analysed bottlenecks in its own inference system and carried out multiple rounds of optimisation. The result was a 31.8% improvement in end-to-end throughput compared with the pre-optimisation baseline — a gain Tencent attributes to the model's own analysis rather than manual engineering work.
Interpreting the 31.8% Figure
A 31.8% throughput gain means the same hardware serves roughly one-third more request volume at equivalent latency. For organisations accessing the model via API, this kind of efficiency gain typically reduces token pricing. For teams running self-hosted deployments, it translates to serving more users from the same cluster, lowering the capital cost of operating a frontier-class model at production scale.
Benchmarks: Where Hy4 Sits
Tencent conducted a blind internal evaluation using 163 domain experts across 203 engineering tasks covering code generation, document analysis, reasoning, and tool use. Hy4 preview scored 2.99 out of 4.00. GLM-5.3 from Z.ai scored 2.92 on the same evaluation, and Kimi K3 from Moonshot AI scored 2.94. The three models are closely competitive — the margin between them is under 0.1 points. Internal evaluations have known limitations because the task set reflects the company's own engineering priorities, and the evaluation was not independently administered. That said, the structured expert-rated methodology across 203 real engineering problems is more contextually grounded than single-benchmark leaderboard comparisons.
Availability and Integration
Hy4 preview weights are available for download from Hugging Face. API access is provided through Tencent Cloud TokenHub and OpenRouter, the latter using an OpenAI-compatible interface that requires no custom SDK integration. Tencent has deployed the model across its own product suite: WorkBuddy and CodeBuddy serve enterprise productivity and code assistance, Yuanbao is a Chinese consumer AI assistant, and ima is a knowledge management application. The split between open weights and managed API access means development teams can choose their deployment model based on cost, control, and infrastructure preference.
What Indian Engineering Teams Should Take From This
For Indian software teams, Hy4 preview offers several practical advantages over comparable commercial alternatives. The Apache 2.0 licence removes the legal ambiguity that more restrictive open-source licences create in commercial deployments — a consideration that matters for enterprise customers in regulated sectors. The 1 million token context window makes the model directly applicable to tasks that have historically required chunking and retrieval pipelines: full codebase comprehension, lengthy contract analysis, and multi-stage agentic workflows that accumulate long reasoning traces. The OpenRouter integration with an OpenAI-compatible interface means teams already using the standard API pattern for ChatGPT or Claude can test Hy4 as a drop-in alternative without changing their application code. For teams evaluating cost efficiency, the MoE architecture prices 49 billion of active compute per token significantly below frontier dense models while retaining frontier-tier quality — a trade-off that is attractive for high-volume production deployments.
The Bottom Line
Tencent open-sourced Hy4 preview on 28 August 2026: a 770B MoE language model with 49B active parameters, a 1 million token context window, and an Apache 2.0 licence. The model improved its own inference stack by 31.8% through autonomous analysis of pipeline bottlenecks. In a blind internal evaluation covering 163 experts and 203 engineering tasks, it scored 2.99 out of 4.00, ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. Weights are on Hugging Face; API access runs through Tencent Cloud TokenHub and OpenRouter. For Indian engineering teams, the Apache 2.0 licence, 1M token context, and OpenAI-compatible API make Hy4 preview one of the more deployable frontier open-weight models currently available.
Frequently Asked Questions
What is Tencent Hy4 preview and when was it released?+
Tencent Hy4 preview is a Mixture-of-Experts language model released and open-sourced by Tencent on 28 August 2026 as the latest in the Hunyuan model family. It has 770 billion total parameters with 49 billion active per token, runs a 1 million token context window, and is licensed under Apache 2.0. The model is available for download from Hugging Face and can be accessed via API through Tencent Cloud TokenHub and OpenRouter using an OpenAI-compatible interface.
How does Tencent Hy4's MoE architecture affect inference cost?+
Hy4 activates only 49 billion of its 770 billion parameters per token — roughly 6.4% of the total weight matrix per forward pass. This reduces inference cost to approximately that of a 49 billion parameter dense model while retaining the representational depth of a 770B system. Additionally, Hy4 autonomously analysed its own inference bottlenecks and improved end-to-end throughput by 31.8%, further reducing the compute cost per token. An FP8 quantised variant also ships alongside the full model for teams running on constrained GPU hardware.
How does Tencent Hy4 compare to GLM-5.3 and Kimi K3?+
In a blind internal evaluation by Tencent using 163 domain experts across 203 engineering tasks, Hy4 preview scored 2.99 out of 4.00. GLM-5.3 from Z.ai scored 2.92 and Kimi K3 from Moonshot AI scored 2.94 on the same evaluation. The three models are competitive — the margin is less than 0.1 points between any two of them. Internal evaluations have limitations because the task set reflects Tencent's engineering priorities, but the expert-rated methodology across 203 real engineering problems is more contextually grounded than single-benchmark leaderboard comparisons.
How can developers access Tencent Hy4?+
Hy4 preview is available as an open-weight download from Hugging Face under the Apache 2.0 licence, making it free for commercial deployment. API access is provided through Tencent Cloud TokenHub and through OpenRouter, which uses an OpenAI-compatible interface requiring no custom SDK. Tencent has also deployed Hy4 in its own products: WorkBuddy and CodeBuddy for enterprise productivity, Yuanbao as a consumer AI assistant, and ima as a knowledge management app. Teams can choose between self-hosted open weights, the FP8 quantised variant for constrained hardware, or managed API access.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.
