Tencent Hunyuan Hy4: 770B, Vision, and Computer Use

What Shipped

Tencent released Hunyuan Hy4 on November 4, 2026, calling it the first Hy model to join the strongest open-weight tier. That phrasing deserves scrutiny — the weights are not out yet — but the release itself is substantial: a 770B-parameter mixture-of-experts model with a 1M-token context, now with vision and computer use.

The pitch is productivity rather than benchmark supremacy. Tencent's own summary is unusually candid: Hy4 is one of the strongest open models, but a clear gap to the top closed models remains. The interesting question is not whether it beats Claude Opus 5.5 — it mostly does not — but whether it is good enough at real work to justify a place in a team's stack at a fraction of frontier prices.

Specs at a Glance

The parameter count is the headline, but the active-parameter figure is what matters for cost: about 49B per token keeps inference affordable despite the 770B total. Tencent says the model was trained partly on multi-token prediction and speculative decoding built into the architecture, which is where the price-performance claim comes from.

Pricing and Access

Pricing is unchanged from the preview, which is the most important commercial fact in the release:

• Input: ¥6 per million tokens
• Output: ¥18 per million tokens
• Cached input: ¥0.3 per million tokens
• Discount: Tencent Cloud TokenHub bills at 50% of list through December 31, 2026
• Free allowance: WorkBuddy/CodeBuddy users get a usage quota from November 4 to 15

Tencent has not published a USD list price. International access runs through OpenRouter and similar gateways, which set their own rates, so compare landed costs rather than assuming a conversion.

Availability is broad for a day-one release: Tencent Cloud TokenHub, WorkBuddy and CodeBuddy, Yuanbao, the WeCom assistant, and ima. More Tencent consumer products are being connected.

Benchmarks: Strong on Real Work, Behind the Frontier

The numbers Tencent chose to lead with are domain-specific, not academic:

• Financial blind test: 213 expert judges — 153 from external financial institutions, 70% with five or more years of experience — scored 579 real tasks. Hy4 averaged 3.04, ahead of DeepSeek V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68). Tencent calls this first among open models. It is a Tencent-designed test, but the external judge pool makes it more credible than a typical vendor chart.
• Office work: average score 83.18 with a 96.15% task pass rate, and first place in 9 of 12 scenarios among domestic models — measured on WorkBuddy workflows, so read it as product-fit evidence rather than a neutral benchmark.
• Game development: +19.1% on GameDev and +9.8% on GameLogic versus the preview, after working with Tencent's own game studios.
• Reliability: internal hallucination rate down from 12.95% to 7.2%, and errors caused by guessing down from 23.1% to 5.6%.

On general benchmarks the picture is more modest. Tencent says Hy4 beats DeepSeek V4.1-Flash-1017 on ProgramBench, FrontierCode 2.0 and Agents' Last Exam v1.1, comes close to Kimi K3.1 Pro on some tests, and exceeds GPT-6.1 Sol on a few — while trailing GPT-6.1 Sol and Claude Opus 5.5 on most. That is the company's own framing, and it is the right one to keep.

What Hy4 Actually Does Better

Three capabilities stand out beyond the benchmark table.

Vision, finally. Hy4 is the first Hy model that can read images: fine-grained photo analysis, chart interpretation, and math problems with diagrams. Tencent's demos — reassembling a 900-piece puzzle into a Monet painting, locating a two-meter rooftop component in a 2.8-gigapixel aerial image — are selected showcases, but they describe a real capability shift for the line.

Computer use aimed at professional software. Most computer-use demos stop at browsers and spreadsheets. Hy4 is trained for Blender, CAD, Unreal Engine, Unity and Godot, with a sandbox of hundreds of applications used for task execution and verification. Whether it survives contact with a real production pipeline is untested by anyone outside Tencent, but the target is the right one.

Office and finance co-writing. The WorkBuddy integration supports human-machine co-writing in Word, Excel and PowerPoint, and Tencent built thousands of financial workspaces and tens of thousands of tasks for domain training. This is the strongest commercial argument for Chinese enterprise buyers: the model is tuned to the software they already use.

The Open-Weights Caveat

"First Hy to join the strongest open models" is a statement about intent, not availability. Tencent says weights will be open-sourced soon, with a system card; the preview shipped under Apache 2.0, but the formal release does not commit to that license.

Until the weights and license are published, treat Hy4 as API-only. If permissive self-hosting is a requirement, Kimi K3.1 Pro (weights due imminently) and DeepSeek's MIT-licensed models are the safer bets today. Watch the license when it lands — Apache 2.0 would make Hy4 one of the more attractive 700B-class options; a restricted community license would change the calculus entirely.

Safety Posture

Tencent describes a deliberately defensive training approach: only defensive cyber data (PoC generation, vulnerability repair) was included, while exploit development, reverse engineering and penetration-testing material was excluded. Audits by Tencent's Xuanwu and Keen labs found no standout offensive capability in cryptanalysis, binary obfuscation or exploitation.

On biological risk, the company reports a weak result on open-ended experiment design — 30.4/100 on its internal Hy-BioExperimentBench — and says it found no evidence the model lowers the barrier to harmful lab work. The caveat is the usual one: these are self-assessments, published before external review.

Limitations

• Weights unavailable. API-only until the promised release; license unknown.
• Vendor benchmarks. The financial and office tests are Tencent-designed, even with external judges.
• Frontier gap. Tencent's own comparison puts most scores below GPT-6.1 Sol and Claude Opus 5.5.
• CNY-only pricing. USD costs depend on gateways and can differ materially.
• China-centric integrations. The strongest workflows assume WorkBuddy, WeCom and Tencent Cloud.
• No independent computer-use evaluation yet. The Blender/CAD claims need third-party testing.

Who Should Use Hy4

• Chinese enterprises already on Tencent Cloud or WeCom — the integration depth is the differentiator, and the price is aggressive.
• Teams doing financial analysis or office document work where the external-judge results are the closest thing to evidence.
• Game and 3D studios using Unreal, Unity or Godot that want an agent tuned for those tools.
• Not yet for teams that need downloadable weights today, permissive licensing, or proven frontier-level reasoning.

Final Verdict

Hy4 is Tencent's most credible model release of the year, and its pitch is unusually honest: strong at real work, still behind the frontier, cheap enough to matter. The 770B MoE, 1M context, vision and computer use put it in the top tier of open-weight candidates — on paper.

Two things decide whether that survives: the license and quality of the promised weights, and whether outside teams can reproduce the Blender/CAD and financial results. Until then, Hy4 is a strong API option for Tencent-centric workflows and a watch item for everyone else.

Related Articles

• GLM-5.3 Review
• Kimi K3.1 Pro Review
• DeepSeek V4.1-Flash Review
• Qwen 3.8-Max Review
• Best Chinese AI Models 2026
• Best AI Coding Tools 2026