Skip to content
Tencent Hunyuan Hy4 Enters Top Open-Source LLMs with Vision, Computer Use & Safety Focus

Tencent Hunyuan Hy4 Enters Top Open-Source LLMs with Vision, Computer Use & Safety Focus

Tencent's Hunyuan Hy4 becomes a leading open-source LLM with visual understanding, computer control, and strong benchmarks in finance, coding, and science, while emphasizing cybersecurity and biosecurity safety audits and maintaining competitive pricing.

Tencent Technical Engineering
Tencent Technical Engineering
Tencent Technical Engineering

Hy4 Overview: Strongest Open-Source Model to Date

Tencent Hunyuan has released Hy4, its latest large language model, which for the first time ranks among the strongest open-source models. Compared to Hy4 preview, Hy4 significantly expands pre-training and post-training data and scale, introduces visual understanding capabilities, and substantially improves instruction following, long-context comprehension, and long-horizon agent performance. Hy4 will soon open-source model weights, publish a model system card, and pursue responsible open-source practices with focus on cyber and biosecurity risks.

Core Capabilities

Visual Understanding, Reasoning & Tool Use

Through visual understanding, reasoning, and tool calling, Hy4 handles fine-grained image processing, complex chart analysis, and mathematical visual reasoning tasks.

Computer Use (Command Line + GUI)

Hy4 supports Computer Use for the first time, combining command-line and graphical interfaces to handle diverse work and life tasks. Specialized training for professional software such as Blender and CAD, as well as cross-application long-horizon tasks, greatly improves planning and execution ability for professional workflows.

Hy4 benchmark comparison chart
Hy4 benchmark comparison chart

Model-Product Co-Design Unlocks Real Productivity Scenarios

Office Productivity: WorkBuddy

Hy4 comprehensively upgrades WorkBuddy user experience. Evaluation shows an average score of 83.18, task pass rate of 96.15%, with 9 out of 12 scenarios ranking first domestically. In practice, Hy4 balances quality and efficiency in visual creativity, 3D scene construction, and web development. For WorkBuddy's human-AI co-writing feature, Hy4 notably improves document aesthetics, stability, and usability.

Long-Term Personal Assistant via Enterprise WeChat Integration

Leveraging improved in-context learning, Hy4 better understands multi-source complex information to serve as a long-term personal assistant. Via WorkBuddy's Enterprise WeChat suite, Hy4 connects chat logs, emails, documents, schedules, and meetings. When multiple projects run in parallel, Hy4 clarifies each project's history and latest progress, identifies pending tasks, required meeting prep, and upcoming deadlines, then reminds users and helps prioritize. Team collaboration features powered by Hy4 will launch with the new "Projects" in WorkBuddy.

Post-trip invoice organization and financial audit
Post-trip invoice organization and financial audit

Professional Financial Analysis

Hy4 strengthens financial analysis capabilities. Using real financial scenarios, thousands of workspaces and tens of thousands of high-quality tasks were built for fund research, securities analysis, quantitative strategy, asset allocation, corporate finance, and tax accounting. Hunyuan partnered with Tencent FiT for crowdsourced testing by 213 experts (153 from top external financial institutions, 70% with 5+ years experience). On 579 real financial tasks, Hy4 ranked first among open-source models (crowdsourced scores: Hy4 3.04, Kimi K3.1 Pro 2.68, GLM-5.3 2.68, DeepSeek-V4.1-Flash-1017 2.97). In internal evaluation by financial AI company Xuntu Tech (IRAB benchmark), Hy4 improved 12.03 percentage points over Hy4 preview.

Search Optimization for Hallucination Reduction

To improve practicality in finance and other office scenarios, Hunyuan launched specialized search optimization for timely retrieval, multi-source integration, and factual consistency verification. In internal tests, Hy4's hallucination rate dropped from 12.95% (preview) to 7.2%, and guess-induced error rate fell from 23.1% to 5.6%.

Frontend & Game Development

For frontend development, Hunyuan optimized complex interactions, business logic, state management, and frontend-backend integration, enabling generated interfaces that are both visually appealing and functionally complete. Collaborating with Tencent IEG professional game developers, Hunyuan continuously improves data and evaluation quality, focusing on game feature implementation, gameplay logic, state understanding, and asset handling. Deep adaptation for mainstream engines UE, Unity, and Godot is achieved, with expansion into multiplayer game development. Compared to Hy4 preview, Hy4 improves 19.1% on GameDev and 9.8% on GameLogic benchmarks.

Retro QQ game arcade demo
Retro QQ game arcade demo

Scientific Research

Hy4 combines complex reasoning, scientific programming, and long-horizon task execution to advance end-to-end research workflows. In the TenureTrail internal benchmark covering 7 disciplines and 123 executable research tasks, Hy4 demonstrates solid research capabilities in mathematical construction, physical modeling, earth environment, systems engineering, and socio-economic decision-making.

Recursive Self-Improvement (RSI) Preparation

Hy4 prepares for the next phase of recursive self-improvement, enhancing model R&D task capabilities. Combined with Hyra2.0, Hy4 optimizes internal pre-training infrastructure for same-architecture smaller models and automatically trains Jev models outperforming official versions on multiple benchmarks. Hunyuan also open-sourced the single-cell representation foundation model HyCell and the physics field foundation model HyPDE, co-developed by Hy4 and Hyra2.0, to convert model-assisted research results into reusable community tools.

Hy4 single-machine 1.5-hour 7M-cell mouse embryo 3D reconstruction
Hy4 single-machine 1.5-hour 7M-cell mouse embryo 3D reconstruction

Responsible Open Source: Cyber & Biosecurity Focus

Cybersecurity

Hy4's training includes only defensive cybersecurity data related to PoC generation and vulnerability patching, while excluding offensive content such as exploitation, reverse engineering, and penetration testing. Security experts from Tencent Xuanwu Lab and Keen Lab audited cryptographic analysis, reverse engineering, and vulnerability exploitation. Results show Hy4 does not exhibit outstanding offensive capabilities in cryptographic reverse solving, anti-binary obfuscation, or vulnerability exploitation.

Biosecurity

Hy4's open-ended experimental design capability lags behind frontier models (Hy-BioExperimentBench: 30.4/100). Internal experts audited trajectories and found strong grasp of basic biological knowledge but limited ability to convert knowledge into executable experimental protocols. No evidence suggests Hy4 significantly lowers the barrier for implementing biochemical harm.

Extreme Cost-Effectiveness: Hy4 Flash Coming Soon

Hy4 is now deployed on Tencent Cloud Tokenhub, WorkBuddy/CodeBuddy, Yuanbao, Enterprise WeChat Smart Assistant Dayuan, ima, and APIs on OpenRouter, Hermes, and OpenCode. Tencent products including QQ, Honor of Kings, Peacekeeper Elite, WeRead, and Tencent News are integrating continuously.

Following a "model practicality" principle, Hunyuan maintains pricing despite large performance gains over Hy4 preview: input 6 CNY/M tokens, output 18 CNY/M tokens, cache hit 0.3 CNY/M tokens. Hy4 flash will launch soon, inheriting Hy4 capabilities while drastically reducing task cost for extreme cost-effectiveness.

To gather feedback, WorkBuddy/CodeBuddy offer free usage quotas from November 4 to 15. Additionally, Hy4 on Tencent Cloud TokenHub is billed at 50% of list price until December 31, 2026.

Since infrastructure rebuild in February, Hunyuan releases a major version every two months on average, first as Preview then refined to official release based on user feedback. Parameter scale continues climbing: Hy3 series under 300B, Hy4 series reaches 770B, and next-gen Hy5 will scale further. The team acknowledges Hy4 still trails top closed-source models and will continue steadily scaling pre-training and post-training to close the gap.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

LLMgame developmentopen-sourcebenchmarkingAI safetyscientific researchfinancial analysisHunyuancomputer usevisual understandingHy4
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.