Tencent Releases Hy4: Vision, Computer Use, and a Planned Open-Weight Drop

At a glance

The release date is November 4, 2026, from Introducing Hy4 and the Transparency Center. The product name is Hy4. The English site uses Tencent Hy; Chinese coverage, including People’s Daily, uses 腾讯混元 Hy4. Versus Hy4 preview, Introducing Hy4 describes larger pre-training and post-training, first vision, first computer use, and gains in instruction following, long context, long-horizon work, and deep research.

OpenRouter’s tencent/hy4 listing gives a context of 1,048,576 tokens and modalities of text + image in → text out. API list price is $0.042 cached / $0.834 input / $2.501 output per 1M tokens, the same as the preview, from Introducing Hy4 and Tencent Cloud. The cloud promotion is 50% of list through 23:59:59 December 31, 2026 (Beijing Time), from the Tencent Cloud pricing page. Open weights are planned soon, with a system card at the weight release, and vLLM and SGLang are named, from Introducing Hy4. Hy4 flash is planned “in the near future,” from Introducing Hy4.

What changed from Hy4 preview

Hy4 preview shipped on August 28, 2026 as a 770B-total / 49B-active MoE with a 1M-token context (official preview post and GitHub model card). The November 4 formal Hy4 post does not restate those parameter counts. People’s Daily, citing Tencent Hunyuan, separately says the Hy4 series reached 770B after Hy3 stayed under 300B—series context, not a fresh first-party architecture dump for the formal checkpoint.

On capabilities, Tencent’s claims are concrete. Hy4 adds vision perception and can combine perception, reasoning, and tool use for fine-grained image work and complex charts. It also adds computer use across CLI and GUI, with targeted training on tools such as Blender and CAD and on multi-application long-horizon workflows. Versus Hy4 preview, Tencent reports +19.1% on GameDev and +9.8% on GameLogic, after co-training with Interactive Entertainment Group developers around Unreal Engine, Unity, and Godot.

In finance and office work, Tencent says it evaluated Hy4 with 213 finance professionals across 579 real-world tasks, ranking it first among open-weight models in that internal panel (Hy4 3.04; DeepSeek-V4.1-Flash-1017 2.97; Kimi K3.1 Pro 2.68; GLM-5.3 2.68). Internal search upgrades cut a reported hallucination rate from 15.76% to 5.45% and speculative errors from 23.1% to 5.6%. On IRAB, a RabyteTech investment-research agent benchmark cited in the post, Hy4 is 12.03 points above Hy4 preview; RabyteTech’s quoted note puts the absolute score at 61.63.

For scientific work, Tencent says Hy4 ranks first among open-weight models on TenureTrail, an internal set of 123 research tasks across seven disciplines. Paired with Hyra 2.0, Hy4 also features in Tencent’s AI R&D narrative; the company says it is open-sourcing two related foundation models, HyCell and HyPDE.

Availability, price, and the open-weight caveat

Hy4 is available now on Tencent Cloud and OpenRouter (tencent/hy4), and in CodeBuddy, WorkBuddy, Yuanbao, and ima. Official list pricing matches Hy4 preview. Tencent Cloud’s pricing page, updated November 4, 2026, adds a limited-time 50% discount through December 31, 2026 (Beijing Time)—a cloud billing note not restated on the Introducing Hy4 research post. OpenRouter currently surfaces that discounted path as $0.417 / $1.251.

Unlike Hy4 preview, which was open-sourced on day one under Apache 2.0, formal Hy4 ships first as a hosted model. Tencent says it plans to open-source the weights, has partnered with vLLM (Inferact Inc.) and SGLang for serving support, and will publish a system card when the weights land. Until then, “among the strongest open-source models” is positioning language for a model whose downloadable weights are still forthcoming.

Safety framing and what Tencent still concedes

Tencent says cybersecurity training data was limited to defensive PoC generation and vulnerability patching, with offensive exploit, reverse-engineering, and penetration-testing material filtered out. Tencent Xuanwu Lab and Tencent Keen Security Lab are named as red-team auditors. On biosecurity, Tencent reports 30.4/100 on internal Hy-BioExperimentBench, below its high-risk threshold, and says it found no evidence that Hy4 meaningfully lowers the barrier to biological harm.

Tencent also concedes a clear gap: even as “one of the strongest open-weight models available today,” Hy4 still trails top proprietary systems on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering. A cheaper, faster Hy4 flash is promised “in the near future.”

Bottom line

Hy4 is Tencent’s November 4 formalization of the Hy4 line: more training, first-party vision and computer-use claims, same API list price as the August preview, and immediate availability across Tencent Cloud, OpenRouter, and several Tencent apps—while the open-weight drop and system card remain scheduled, not shipped. Treat the finance, agent, and science scoreboards as Tencent (and partner) evaluations unless independent leaderboards catch up; treat “open-source / open-weight” language as intent plus preview lineage, not as proof that Hy4 weights are downloadable today.