Aug 29, 2026

Tencent Hy4 preview: The Open Model Stretching Context to 1M

Back to blog

One million tokens has stopped sounding like a lab demo and started looking like a product spec. On August 28, 2026, Tencent released and open-weighted Tencent Hy4 preview, a Mixture-of-Experts model with 770 billion total parameters, 49 billion active per token, and a context window above 1M tokens.

The official Tencent announcement frames Hy4 preview as a tool for real productivity: coding, office work, financial analysis, game development, and scientific research. TechNode confirms the timing and summarizes the bet: open weights, integration into WorkBuddy and CodeBuddy, APIs through Tencent Cloud TokenHub and OpenRouter, and an attempt to compete at the open AI frontier without relying only on academic benchmarks.

Official Tencent Hy4 preview launch image
Hy4 preview arrives as the new flagship model in the Tencent Hy family, focused on long productivity tasks. Image: Tencent

From Hy3 to Hy4: a short timeline

  1. April 2026: the Hy3 preview family put Tencent into the race for open models aimed at agents and programming.
  2. July 2026: the company introduced more product-integrated Hy3 releases tied to CodeBuddy and WorkBuddy, using real feedback to tune behavior.
  3. August 28, 2026: Hy4 preview appears with larger scale, 1M context, variants published on Hugging Face, ModelScope, GitCode, and CNB, plus code and documentation on GitHub.
  4. September 2026: Tencent promises temporary free access in WorkBuddy and CodeBuddy while the Hy4 generation continues under a preview-first approach.

What separates this from a routine model launch is the mix of scale and packaging. The repository and model page describe a MoE architecture with 78 layers, 256 routed experts per layer, one shared expert, and eight routed experts activated per token. Instead of running everything for every request, the system tries to spend compute where it matters. That is the modern way to chase giant models without turning every call into a cost fire.

The 1M-token context is the part that changes the conversation for software teams. An agent reading several files, logs, tickets, documentation, and decision history often fails less because it lacks abstract intelligence and more because it loses the thread of the job. A much larger window does not solve memory, permissions, or verification, but it reduces the need to chop a problem into artificial pieces.

Official Tencent visual about Hy4 preview capabilities
Tencent positions Hy4 preview for programming, analysis, office workflows, games, and scientific research. Image: Tencent

The number that impresses and the number that matters

770B parameters is impressive. But the number that may matter more day to day is 49B active parameters. That partial activation helps determine whether a model can be served competitively. Tencent lists API pricing at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-hit tokens. The OpenRouter page already lists the model as available through Tencent Cloud, with 1M context and an August 28 release date.

There is also an interesting evaluation figure: in an internal blind assessment with 163 experts and 203 engineering tasks, Tencent says Hy4 preview averaged 2.99 out of 4, slightly ahead of GLM-5.3 and Kimi K3. That is useful, but it should be read carefully. Internal evaluations explain product intent; they do not replace independent tests, real workloads, or total-cost comparisons.

The loop that is starting to close

The most interesting detail in the announcement is not only that the model writes code. Tencent says Hy4 preview participated in its own development process: training optimization, data strategies, evaluation frameworks, low-level operators, and inference bottlenecks. The cited result is a 31.8% increase in end-to-end throughput versus a baseline.

That does not mean a model magically improves itself. It means something more concrete and perhaps more important: AI labs are building loops where models help improve tools, tools accelerate new models, and products supply usage feedback. For people who build software, that loop is starting to look as relevant as the model architecture itself.

Official Tencent Hy4 preview image about AI productivity
The launch emphasizes real product feedback as part of the Hy4 evolution cycle. Image: Tencent

Where caution still belongs

The Hugging Face model page itself calls this an early version. Known limitations include spending longer than necessary reasoning through complex tasks and over-verifying its own work. In coding agents, both flaws can cost money, time, and patience: a model that thinks too much can be expensive; a model that verifies too much can block workflows that needed a simple decision.

For technical communities, small teams, and independent builders, the practical conclusion is balanced. Hy4 preview matters because it combines open weights, huge context, an Apache 2.0 license, deployment recipes for vLLM and SGLang, and API access. But it should still be treated as a preview: test it on your own tasks, measure latency, check cache-driven costs, observe tool calling, and keep humans in the loop for decisions that touch production.

The launch matters now because it shows a new stage in the open-model race: publishing a big model is no longer enough. It needs long context, understandable pricing, deployment paths, product integration, and a convincing story about how it improves. Hy4 preview may not settle that race, but it forces everyone else to answer at a longer scale.

Comments (54)

Sign in to comment. Register