Models#Benchmark#Device control
Tencent ships Hunyuan Hy4: vision, computer use, and open weights soon
Tencent shipped Hunyuan Hy4 on November 4: its first model with vision and computer use, hallucination down to 5.45% per Tencent, open weights promised.
Tencent released the final version of Hunyuan Hy4 on November 4 in Beijing, about ten weeks after the August 28 preview. Hy4 is a 770B-parameter mixture-of-experts model and the first in the Hunyuan line that reads images and operates a computer, driving graphical interfaces and the command line within one task. Tencent says a self-built search test cut the hallucination rate from the preview’s 15.76% to 5.45%. API prices stay at preview levels — 6 and 18 yuan per million input and output tokens, roughly $0.85 and $2.50 — and the weights are promised “soon”.
Key points
- New capabilities: Hy4 is the first Hy model with vision (images, complex charts, math-visual reasoning) and computer use, mixing GUI and terminal in one session; Tencent says it trained specifically on Blender, CAD, UE, Unity and Godot workflows
- Scores: 3.04 in a crowd test where 213 finance experts set 579 real tasks, ahead of DeepSeek-V4.1-Flash-1017 (2.97) and Kimi K3.1 Pro and GLM-5.3 (2.68 each) — all per Tencent’s own evaluation
- Open weights and price: Tencent calls Hy4 a first-tier open model and promises weights plus system cards soon (the preview has sat on Hugging Face under Apache 2.0 since August); API pricing is unchanged and TokenHub tokens are half price through December 31
Background
The Hy4 preview shipped on August 28 with its weights already public: 770B total parameters, 49B active per token, 78 layers, 256 routed experts plus one shared expert with top-8 routing, a 1M-token context, all under Apache 2.0 on Hugging Face. That model card claimed software engineering, office documents, game development and research — no vision input, no computer control. Tencent’s news channel fills in the cadence: the company rebuilt its training infrastructure in February and has shipped a major version roughly every two months since. The final Hy4 enlarges pre-training and post-training scale and bolts on the two missing capability families. Release the API first, follow with weights is also the pattern Chinese labs now share — MiniMax ran the same play with M3.1 Flash in October.
The facts
Every number below is a vendor claim or a paraphrase of Tencent’s announcement; no independent retest exists yet.
| Item | Hy4 final release (official figures) |
|---|---|
| Released | November 4, 2026, Beijing |
| Architecture | 770B-total MoE (per Tencent News); preview ran 49B active, 1M context |
| New capabilities | Vision understanding; computer use (GUI plus command line) |
| Finance crowd test | 3.04 across 213 experts and 579 real tasks |
| Same-table rivals | DeepSeek-V4.1-Flash-1017 2.97; Kimi K3.1 Pro and GLM-5.3 2.68 |
| Hallucination rate | 15.76% to 5.45% on a self-built search test (per ifeng’s copy of the release); Tencent News’ tech desk prints 12.95% to 7.2% |
| Guessing-error rate | 23.1% to 5.6% (both channels agree) |
| Office tasks | WorkBuddy average 83.18, 96.15% pass rate, best domestic score in 9 of 12 scenarios |
| Games and biology | GameDev +19.1%, GameLogic +9.8% vs preview; Hy-BioExperimentBench 30.4/100 |
| Pricing | 6 / 18 yuan per million input / output tokens, 0.3 yuan cache hits — same as preview |
| Open source | Weights and model system cards “coming soon”; the preview is Apache 2.0 |
Three details deserve a second look. The finance-test rival — the same DeepSeek-V4.1-Flash-1017 that Unsloth just quantized down to 64GB machines — trails by 0.07 points, hardly a class difference. The hallucination figures disagree across channels: ifeng’s copy of the announcement says 15.76% fell to 5.45%, Tencent News’ tech desk says 12.95% fell to 7.2%, most likely two different self-built test sets, and neither has been checked by anyone outside Tencent. The office and gaming gains (83.18, 19.1%) rest solely on Tencent’s own benchmarks too.
What others say
The sharpest commentary so far comes from Tencent itself. The official statement, as quoted by tech outlet IThome, admits Hy4 still lags top closed models by a wide margin on hard sets for math, long-horizon development, model iteration and visual question answering. The security write-up is specific: cyber training data was limited to defensive material (PoC generation, vulnerability fixing) with offensive content stripped, and auditors from Tencent’s Xuanwu and Keen labs found no prominent offensive capability. On biology, the model scores 30.4 out of 100 on Tencent’s own experiment-design benchmark — strong domain knowledge, weak conversion into executable lab protocols. Ecosystem support is lined up through the vLLM and SGLang communities and Inferact Inc., with listings on OpenRouter, Hermes and OpenCode; the model is rolling into Tencent’s Yuanbot assistant, ima and enterprise WeChat, with QQ and Honor of Kings to follow. WorkBuddy and CodeBuddy carry free quotas from November 4 to 15. As of publication, no independent benchmark has retested the final model — reason enough to read every score above as “Tencent says”.
Our take
For agent builders, Hy4 fills a specific gap: no open-weight model has yet combined vision, computer use and a 1M-token context, and 6/18-yuan pricing undercuts OpenAI’s GPT-6.1 Terra ($1/$5 per million tokens), the cheapest closed computer-use option we covered in October. Distribution through enterprise WeChat, Yuanbot and ima is a ready-made channel for products selling into Chinese companies. Two things should settle before anyone commits production budget: the final weights’ actual license and size — the preview is Apache 2.0, the final release publishes nothing yet — and an independent retest, because when a vendor’s own press copies disagree about its headline hallucination number, that number is a claim, not a fact. Credit where due: Tencent’s own admissions (biology at 30.4/100, real gaps on hard math and long-horizon coding) are the honest part of this launch. A cheaper, faster Hy4 Flash and a larger Hy5 are already on the roadmap, and a model every two months is a pace few teams can keep up with.