Tencent Hy4 Preview: Inside the New 770B Open-Weight Model (2026)

Tencent open-sourced the Tencent Hy4 preview on 28 August 2026 — a next-generation Hunyuan model with 770 billion total parameters, 49 billion active per token and a context window beyond one million tokens — and, per the launch coverage, you can use it free for a limited two-week period inside WorkBuddy and CodeBuddy. That is the headline: a frontier-scale open-weight model, priced aggressively on the API, that Tencent says even helped optimise its own training pipeline. In this guide I will break down what shipped, what the benchmark numbers actually say, how to get access today, and where it fits in an AI business stack.

📺 Watch: NEW Tencent Hy4 is Mind Blowing

🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside

I have been covering the Chinese open-weight wave closely — my Tencent HY3 review from the previous generation is one of the most-read pages on this site — and Hy4 is the most ambitious release Tencent has put out yet. The previous HY3 was a 295-billion-parameter mixture-of-experts model with 21 billion active parameters, built for agent workloads rather than frontier coding. Hy4 more than doubles the total parameter count and pushes hard into exactly the territory HY3 fell short in.

What the Tencent Hy4 Preview Actually Is

According to Tencent's announcement and the reporting around it, the Hy4 preview is a mixture-of-experts model with 770 billion total parameters, of which 49 billion activate per token. That sparse design matters for practical use: you get frontier-scale capability while the compute cost per token stays closer to a mid-sized model. The context window exceeds one million tokens, which puts entire codebases, long document sets and multi-day agent sessions inside a single context.

Tencent positions the model for productivity work rather than chat: coding, office tasks, data analysis, game development and scientific research are the use cases named in the announcement. That framing lines up with where the benchmark gains landed, which is worth looking at properly.

The Benchmark Story: A Generational Jump on Agentic Coding

Per the launch reporting, Hy4 preview achieved a generational leap across twelve benchmark tests. Two numbers stand out. On Terminal Bench 2.1 — the terminal-driven agentic coding benchmark — it scored 85.4, which the coverage notes puts it above DeepSeek V4 Pro. And on DeepSWE, the software-engineering benchmark, the Hunyuan line jumped from 28.0 to 64.3 generation-on-generation. Those are exactly the benchmarks that predict how well a model runs inside an agent harness, which is why I pay more attention to them than to trivia-style leaderboards.

Tencent also ran an internal evaluation with 163 experts across 203 real engineering tasks. The reported averages, best score first:

Treat vendor-run evaluations with the usual caution — every lab's internal numbers flatter its own model — but the external Terminal Bench result points the same direction, and that consistency is what makes me take the Tencent Hy4 preview seriously rather than filing it as another headline parameter count. For comparison context on how the current frontier models stack up against each other, my DeepSeek V4 Pro vs Claude Fable 5 vs Grok 4.6 breakdown covers the models Hy4 is now gunning for.

The Model That Helped Train Itself

The most interesting claim in the announcement is not a benchmark at all. Tencent says Hy4 participated in optimising its own training pipeline — the model was used during its own development to help automate optimisation of training methods — and separately optimised its own inference infrastructure, delivering a measured 31.8 per cent end-to-end throughput increase against the baseline. In plain terms: the model made itself cheaper and faster to run before it ever shipped.

Whether you find that exciting or slightly unnerving, it is a preview of where the whole field is heading. Models that improve the systems around themselves compound in a way that raw parameter counts never did, and it is precisely the kind of loop that makes agentic AI worth building a business on now rather than waiting another cycle.

Want every major model drop like Hy4 tested, compared and wired into working business workflows within days — instead of doing the digging yourself? That is exactly what we do inside the AI Profit Boardroom → Get the breakdowns as they land

📺 Watch: Build Anything with Tencent Hy4!

How to Access the Hy4 Preview Today

Availability at launch is broader than most Chinese releases manage. Per the announcement, the options ranked by how quickly you can be using it:

1. WorkBuddy and CodeBuddy — free for two weeks. Both the Chinese and international versions offer complimentary Hy4 access for a limited two-week period. If you just want to feel the model out on real tasks, start here; it costs nothing during the window.
2. Yuanbao and ima. Tencent's consumer-facing products also carry the model at launch.
3. API via Tencent Cloud TokenHub and OpenRouter. For agent work this is the route that matters. Introductory API pricing is 0.834 US dollars per million input tokens and 2.501 dollars per million output tokens — aggressive for a 770B-class model. If you already route models through OpenRouter, it drops into your existing setup; my guide to running Hermes agent through OpenRouter covers exactly how that wiring works.
4. Open weights. The preview is open-sourced, so self-hosting is on the table for teams with the hardware — though at 770 billion total parameters this is not a laptop model, and most people will be better served by the API.

Where Hy4 Fits in an AI Money Stack

Here is my practical read for anyone building revenue with agents rather than collecting benchmarks. The million-token context plus the Terminal Bench score make the Hy4 preview a genuine candidate for the heavy lifting inside agent harnesses: long research runs, repository-scale coding tasks, and document-heavy workflows like proposals, audits and data analysis. The API pricing means you can experiment without the bill dictating your architecture.

We run every notable model through Goldie Bench, our own benchmark, before it earns a slot in any client-facing workflow, and the pattern our testing keeps confirming is that agentic benchmark gains translate to real-world agent reliability far better than chat-style scores do. Hy4's profile — strong Terminal Bench, huge context, cheap input tokens — is the profile that tends to do well on that kind of testing. Once a model earns its place, it slots into the Agent OS, our system for turning raw models into working business automations, as a routable option alongside the incumbents rather than a wholesale replacement.

If you are choosing a daily driver for a Hermes-style setup today, my best model for Hermes agent guide is the fuller decision framework — Hy4 has to prove itself on reliability and latency before it displaces anything, and two weeks of free WorkBuddy access is the cheap way to find out.

Hy4 Preview vs HY3 and the Chinese Open-Weight Wave

Against its predecessor, the jump is stark: HY3 was 295 billion total parameters with 21 billion active and honest-but-modest coding ability, while the Hy4 preview lands at 770 billion with 49 billion active and benchmark results that compete at the top of the open-weight table. My HY3 review still applies if you want the budget agent-task option, but Hy4 is playing a different game.

The wider context matters too. Hy4 arrives days after Z.ai shipped GLM 5.3 Flash and weeks after Moonshot's Kimi K3 open-weight release — which you can still use free — and the pace is now relentless. My overview of the Chinese AI model landscape covers why these labs keep open-sourcing frontier-scale work, but the practical takeaway for business builders is simple: capable model access keeps getting cheaper, and the advantage shifts to whoever has the systems to deploy it fastest.

Quick Reference: Tencent Hy4 Preview

Tencent Hy4 Preview FAQs

Is the Tencent Hy4 preview free to use?

Partly. WorkBuddy and CodeBuddy users get complimentary access for a limited two-week period, and the preview weights are open-sourced. Beyond that, API access is paid — 0.834 dollars per million input tokens and 2.501 dollars per million output tokens at the introductory rate.

How big is the Hy4 context window?

Over one million tokens, per Tencent's announcement — enough for repository-scale codebases or very long agent sessions in a single context.

Is Hy4 better than DeepSeek V4 Pro?

On Terminal Bench 2.1 specifically, the reported 85.4 score surpasses DeepSeek V4 Pro. Overall superiority is a bigger claim than one benchmark supports — run your own tasks through the free WorkBuddy window before switching anything that earns money.

What did Hy4 optimise about its own training?

Per the announcement, the model was used during its own development to help automate optimisation of the training pipeline, and it separately optimised its inference infrastructure to a measured 31.8 per cent throughput increase.

Verdict: The Most Serious Open-Weight Release of the Month

The Tencent Hy4 preview is the strongest signal yet that open-weight models are competing at the top of the agentic coding table rather than trailing it. The specs are frontier-scale, the agentic benchmarks are the right ones to be good at, the pricing undercuts the incumbents, and there is a genuinely free two-week window to test it yourself. My advice is what it always is: ignore the hype cycle, run it on your own real tasks, and let results — not launch posts — decide whether it enters your stack.

If you would rather skip straight to what works — every new model tested for you, plus the exact agent workflows we run on whichever one wins — join us inside the AI Profit Boardroom → Turn model drops into money