Tencent's Hy4: A Free Rival to GPT and Claude, With a Catch

Tencent just open-weighted a frontier AI model that, by its own admission, still loses to Claude and GPT on the hardest problems — and the gap is closing fast enough that nobody in the AI world is laughing it off anymore.

A Chinese model built to be copied

On November 4, Tencent announced Hy4, the newest version of its Hunyuan model family, and positioned it as one of the world's leading open-source AI systems. Under the hood, Hy4 uses a mixture-of-experts (MoE) design — instead of running every one of its parameters for every single request, it has 770 billion parameters total but only activates around 49 billion of them per answer, which keeps it faster and cheaper to run than a dense model of similar size. It also reads up to a million tokens of context in one go (roughly the length of a thick novel), and the full release adds visual understanding and "computer use," meaning the model can look at a screen and operate software — clicking buttons, filling forms — the way a person would.

How close is close?

Tencent published its own comparison tables against the earlier Hy3 model, and the jump is real: scores on SWE-bench Multilingual (a test of fixing real GitHub issues across languages) went from 75.8 to 82.9, SWE-bench Pro climbed from 57.9 to 65.7, and DeepSWE more than doubled, from 28.0 to 64.3. Reporting around the launch put Hy4 in direct comparison with:

• Claude Opus 5 and GPT-5.6 Sol (closed, proprietary models)
• DeepSeek V4 Pro, GLM 5.3, and Kimi K3 (other Chinese open-weight rivals)

Tencent itself was unusually candid about where it still falls short, acknowledging a real gap against the top closed-source models on its hardest internal benchmarks, research-grade mathematics, and long-horizon coding tasks that require staying coherent across many steps.

Catching up to the frontier used to take a well-funded lab years. Now it looks like it takes one earnings quarter.

"Open" comes with an asterisk, for now

Here's the catch worth sitting with: at the moment Tencent made its announcement, the actual model weights — the trained numerical parameters you'd need to download and run Hy4 yourself — weren't available yet. Coverage from launch day found the official Hugging Face page returning an authorization error, with Tencent saying weights and a system card would follow "in the coming days." An earlier preview version of Hy4, released back in August under the permissive Apache 2.0 license, is already downloadable, so the pattern — preview first, polished flagship release second, weights trickling out after the press cycle — is becoming Tencent's playbook rather than a one-off delay.

What this means if you're building AI in India

For Indian developers and startups, the economics here matter more than the bragging rights. Running inference against a closed API from OpenAI, Anthropic, or Google means paying in dollars for every token, which adds up fast for a team billing customers in rupees. An open-weight model you can self-host, even one this large, changes that math — it's part of why the IndiaAI Mission has been pushing for shared GPU compute capacity and why homegrown efforts like Sarvam AI and Krutrim have leaned on open-weight foundations rather than building everything from scratch.

But there's a second layer Indian teams can't skip past: Hy4 is a Chinese company's model, built and trained inside China's regulatory environment. India has a recent, concrete history of restricting Chinese-origin software over data and security concerns — the 2020 ban on more than 200 Chinese apps is the obvious precedent — and government bodies and regulated sectors like banking will likely want their own security review before any Hy4-based tooling touches production systems, even when the weights are sitting locally on their own servers and nothing is calling home to a Chinese cloud. Open weights remove the pricing objection; they don't automatically remove the trust question.

The race just got another lane

What Hy4 really signals is that the frontier AI race isn't a three-lab story anymore. A model that was trailing by a wide margin six months ago is now close enough to force real decisions — not just "which model is smartest," but "which trade-off between cost, capability, and who built it" a team is willing to live with. For Indian developers watching their API bills, that's a genuinely useful new option. For anyone deciding whether to put it into production, it's homework, not a free lunch.