Tencent's Hy4 adds vision and computer control, weights still closed

Summary

• Tencent released the full Hy4 model with image input and computer-use capability on November 4.
• The August preview was open-sourced under Apache 2.0; the full version's parameter count and license are unannounced.
• Pricing stays at preview levels, putting an agentic Chinese model into the same tier as cheaper Western offerings.

The latest

Tencent moved its Hy4 model out of preview on November 4, 2026, adding visual input and computer use for the first time in the Hy line. The company says the model reads fine detail in images and complex charts and operates both command-line and graphical interfaces, after targeted training that covered Blender, CAD software and multi-app workflows. Prices were left unchanged.

Details

• What's new: Tencent says Hy4 combines reasoning, tool calling and long-horizon agentic work, released in Chinese and English. Vision and computer use are new to the Hy family, and the company attributes interface control to training on desktop applications and workflows spanning several programs in sequence.
• The preview baseline: Hy4 Preview shipped on August 28, 2026 as open source: 770 billion total parameters, 49 billion active, a 1 million-token context window and an Apache 2.0 license. Tencent acknowledged then that the preview over-reasoned and over-verified. It has not stated the full version's parameter count or license.
• The claimed gains: Tencent says it substantially expanded pre-training and post-training, reporting improvements in instruction following, long-context understanding, long-horizon agentic tasks and deep research, with no figures attached. It says the over-reasoning and over-verification problem has been reduced.
• Internal numbers: On a financial evaluation of 579 tasks judged by 213 finance professionals, 153 of them from financial institutions, Tencent reports Hy4 scoring 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3, describing it as first among open-weight models.
• Error rates: Tencent says hallucination in internal research evaluations fell from 15.76% in the preview to 5.45%, and speculative errors from 23.1% to 5.6%. On TenureTrail, 123 research tasks across seven disciplines, it claims first place among open models; on Hy-BioExperimentBench it reports 30.4 out of 100.
• Gaming and partners: Work with Tencent's interactive entertainment developers yielded a claimed 19.1% improvement on GameDev and 9.8% on GameLogic over the preview, covering Unreal Engine, Unity and Godot. Digital Applied cites partner results: Manus near 97% completion on document tasks, RabyteTech 61.63 on IRAB, Yuanbao 72.53 against 65.25.
• Admitted gaps: Tencent concedes Hy4 trails the best closed models in research mathematics, long-horizon coding, model self-improvement and advanced visual question answering. It says it used its Hyra 2.0 autonomous research agent to improve the pre-training architecture of smaller models in the family, calling that a next stage of iterative self-improvement.
• Pricing: List price per million tokens is $0.834 input, $2.501 output and $0.042 cached input, unchanged from the preview; the Chinese posting lists 6, 18 and 0.3 yuan. SiliconANGLE rounds these to roughly 90 cents, $2.70 and 4.5 cents. OpenRouter advertises a 50% discounted route at $0.417 and $1.2505 with no stated expiry.
• Availability: Access runs through Tencent Cloud's TokenHub, OpenRouter, Hermes, OpenCode and the CodeBuddy, WorkBuddy, Yuanbao, WeCom assistant and ima apps. OpenRouter lists a 1,048,576-token context, a 65,536-token output ceiling, text and image input, and Tencent Cloud as sole provider.
• The weights: Tencent says "We plan to open-source Hy4's model weights" alongside a system card, with vLLM and SGLang community support, and has set no date. Digital Applied reported that on November 5 the tencent/Hy4 page on Hugging Face returned 401 and the GitHub repository 404. A cheaper Hy4 Flash is promised without a date.
• Safety review: Tencent says cybersecurity training data was limited to defensive proof-of-concept generation and vulnerability patching, with exploit development, reverse engineering and penetration testing filtered out. Red teams from Tencent Xuanwu Lab and Keen Security Lab found no high risk in cryptographic reversal, binary deobfuscation or exploitation. Internal reviewers judged the model's ability to turn biological knowledge into executable protocols limited, deferring full evaluation to the unscheduled system card.

Between the lines

The gap between the two releases is the story. The preview came with a parameter count, a context window and an Apache 2.0 license; the full version arrives with benchmark claims Tencent generated itself and no license at all. With Hugging Face and GitHub both unreachable, the open-weight positioning Tencent uses in its own financial-benchmark framing currently rests on a plan rather than a download.

What's next

Watch for the Hy4 system card and the weights release, both undated, the Hy4 Flash launch, and whether OpenRouter's 50% discount route persists. Independent benchmark results would be the first outside test of Tencent's hallucination and error-rate claims.

Source: Unite.AI, SiliconANGLE, Digital Applied, Tencent