Tencent releases Hy4 Preview, a 770B-parameter open model with a 1M context window, and says it narrowly beat GLM-5.3 and Kimi K3 in internal engineering tests
Bloomberg
Context & Ripple Effects
Tencent had already moved Hy3 from an April preview into a July Apache 2.0 release, positioning the 295B-parameter model against GLM-5.1 and 5.2. Hy4 is a far larger follow-on release, pairing its claimed engineering-test edge with a 1M-token context window.
The launch lands amid an escalating Chinese open-model race: Alibaba’s 2.4T-parameter Qwen3.8 Max preview was also pitched against frontier systems, while Tencent’s July Hy3 release emphasized competitive performance at a smaller scale. Public reactions focused on Hy4’s coding-agent potential, though Tencent’s comparison with GLM-5.3 and Kimi K3 is based on its internal tests.
First-order effects
Developers and Tencent’s product channels gain access to a 770B open model designed for coding, research and other long-context workloads, expanding Tencent’s foundation-model offering beyond Hy3.
GLM-5.3 and Kimi K3 face a new directly positioned open-model rival; Tencent’s claimed advantage is limited to its internal engineering evaluation.
Second-order effects
Z.ai and Kimi must compete not only on model quality but on the serving stacks and developer integrations already built around their models, a switching-cost concern raised in public discussion.
A 1M-token context window makes long-document and agentic coding workflows a more explicit competitive dimension, increasing the value of infrastructure that can serve such workloads efficiently.
Third-order effects
If successive releases keep combining larger open weights with frontier-oriented coding claims, model competition shifts from isolated benchmark wins toward the availability, inference cost and integration of deployable systems.
The pattern points to open-weight models serving as an entry point while product distribution channels such as Tencent’s WorkBuddy and CodeBuddy determine which models reach working users.
The trend:Chinese AI labs are pairing rapidly scaled open models with long-context and coding capabilities, making distribution and serving economics as consequential as headline benchmark claims.
Big news: Hy4 preview by @TencentHunyuan just landed ~#5 in the Code Arena: WebDev with 1633 pts (AutoEval). This is a significant improvement from Hy3 at #31 overall (+115 pts)! Among open models, Hy4 preview is ~#3, compared to Hy3 at #7. Note: this is an early AutoEval
New Hy4 model release from tencent is actually pretty good. GLM-5.3 size model at the same coding level. Won't be too popular I suspect since people have already optimized serving for GLM5.2 and all of that will translate over to 5.3.
Tencent Just released a GLM-5.3 competitor, 740B params, reaches at top 15 models Very nice pace of iteration for Chinese labs, they are in flow. I've tested Hy3, very solid. https://huggingface.co/...
Built for real-world productivity. Hy4 preview delivers strong performance across coding, office work, game development and scientific research — now available globally through WorkBuddy, CodeBuddy, Yuanbao and more. Try Hy4 preview. OpenRouter: https://openrouter.ai/... Tencent
I gave WorkBuddy one prompt: build a campaign dashboard I could actually use. It planned the app, created CreatorFlow, ran it locally, and tested it on my computer. The result is a working system for campaigns, approvals, deliverables, and payments. Built with Hy4 Preview.
Chinese AI models are advancing relentlessly! Tencent's Hy4 preview is another near-frontier open-source model. It seems at the Kimi K3 level while being 3 times smaller at 770B parameters! In particular, I am interested in the high scientific reasoning benchmarks! Looking good!
Hy4 Preview is out! 770B-A49B, 1M context, Apache 2.0 Reported scores are great, putting it roughly on par with GLM-5.3 (!). Tencent has REALLY picked up model training since @ShunyuYao12 joined, congrats!
Tencent just released Hy4 Preview: a 770B-parameter open-source model with 49B active parameters and a 1M-token context window. The most interesting part is its agentic research capability. Tencent says Hy4 coordinated several Codex sessions in parallel, evaluated their results
🚀 Hy4 preview is here. 770B, 49B active, 1M context. Built for productivity. Open source frontier. Consistent affordable price. Use it. Tell us what breaks. More on Hy blog： https://hy.tencent.ai/... HuggingFace： https://huggingface.co/... Github：https://github.com/ ...
Tencent says it's a early version, and that they will continue to polish it post release, but it already appears to be very strong. They have been increasing their AI spend, looks like it's starting to pay off. Input: $0.834/m Output: $2.501/m
Wow. Given Hy3-preview to Hy3 trajectory, I think Hy4 should exceed Opus 5. Another derivative of DeepSeek architecture, which makes it lowkey tragic how it dunks on the Whale-0813 in... almost everything, at ≈GLM-5 size. Tencent wants to join the top tier.
Hy4-Preview is now on @OpenRouter @ $2.5/1M. 👀 This is reasonable considering it's benchmarking right alongside Kimi K3 (@ $12.75/1M) and GLM-5.3 (@ $1.2/1M) Let's test it.
...The wildest part. It matches DeepSeek V4 Pro's active parameter count at roughly half the total model size. And scores higher. Their last model Hy3 had 295B total and 21B active. This is a 2.5x jump in one generation. Open source is not catching up to frontier anymore. I…
Tencent has released and open-sourced Hy4 preview, a next-generation large language model with 770B total parameters and a 1M+ token context window. Ranked among the top tier of open-source models. Try Hy4 preview. OpenRouter: https://openrouter.ai/... Tencent Cloud TokenHub:
@TencentHunyuan Shipping it live in @CommandCodeAI 🐐 Excited to be an early partner to benchmark Hy4 preview, in our internal bench it performed better than GLM 5.3 Flash on token costs and time. Will continue to test and share more results.
Introducing Hy4 preview ◆ Open weights: 770B MoE, 49B active, 1M context ◆ Built for real work: code, docs and analysis, scientific research ◆ Hy4 preview helped optimize its own training pipeline, kernels included ◆ We ship upgrades, not price hikes Free in @WorkBuddy_AI