Hy4 API

Hy4 is the official release of Tencent Hy's latest-generation large language model, announced on November 4, 2026 as the successor to Hy4 preview. Tencent says Hy4 places Hy among the strongest open-source models for the first time. Compared with Hy4 preview it substantially scales up both pre-training and post-training, adds vision for the first time and delivers major gains in instruction following, long-context understanding, long-horizon tasks and deep research.

Tencent has announced that Hy4's model weights and system cards will follow, with its open release focused on responsible handling of cybersecurity and biological safety risks. Until then, Tencent's launch post, press release and benchmark appendix are the official sources for this card. Tencent has not yet published Hy4's architecture details. Hy4 preview, the release it builds on, is a 770B-parameter Mixture-of-Experts model with 49B parameters active per token.

Key capabilities reported by Tencent:

• Vision. Hy4 supports visual perception for the first time. By combining perception, reasoning and tool use it can understand fine-grained image details and interpret complex charts. It also handles mathematical visual reasoning.
• Computer use. Hy4 also supports computer use for the first time, working across command-line interfaces (CLI) and graphical user interfaces (GUI). Tencent trained it specifically on professional software such as Blender and CAD and on long-horizon workflows that span multiple applications, significantly improving its ability to plan and carry out professional tasks.
• Front-end development. Tencent optimized Hy4 for complex UI interactions, business logic, state management and full-stack integration, so generated interfaces are both visually polished and functionally complete.
• Game development. Working with professional developers at Tencent Interactive Entertainment Group (IEG), Tencent focused training and evaluation on feature implementation, gameplay logic, state tracking and asset pipelines. Hy4 is adapted for Unreal Engine, Unity and Godot and extends to multiplayer game development. Compared with Hy4 preview it improves by 19.1% on GameDev and 9.8% on GameLogic.
• Finance and office workflows. Tencent built thousands of realistic workspaces and tens of thousands of tasks spanning fund research, equity analysis, quantitative strategies, asset allocation, corporate finance and tax accounting. With Tencent Financial Technology (FiT), 213 finance professionals (153 of them experts from leading financial institutions, more than 70% with over five years of experience) evaluated 579 real-world financial tasks. Hy4 ranked first among open-weight models with a score of 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68). On IRAB, an investment-research benchmark from RabyteTech, Hy4 scores 12.03 points higher than Hy4 preview.
• Fewer hallucinations. Targeted search upgrades focus on time-sensitive retrieval, multi-source synthesis and factual consistency checks. In Tencent's internal evaluations the hallucination rate fell from 15.76% on Hy4 preview to 5.45% and speculative errors fell from 23.1% to 5.6%.
• Working with documents. Hy4 can co-edit Word, Excel and PowerPoint files with people in real time. In blind evaluations Tencent reports clear gains over Hy4 preview in document formatting, reliability and overall usability. On Tencent's WorkBuddy evaluations Hy4 averages 83.18 with a 96.15% pass rate and ranks first in nine of twelve scenarios in the mainland China market.
• Research and self-improvement. Tencent is preparing Hy4 for recursive self-improvement. With the Hyra 2.0 harness, Hy4 optimized the pre-training infrastructure for smaller models that share its architecture. The Hy team also open-sourced two scientific foundation models developed with Hy4 and Hyra 2.0: HyCell for single-cell representations and HyPDE for physics fields.

Benchmarks

Reproduced from the benchmark appendix in Tencent's official Hy4 launch post. Each model is evaluated at its highest available reasoning setting. In the cells, * marks Tencent's own testing and † marks evaluation on the text-only subset of HLE. Where a cell holds two values both are as published. Benchmarks marked ‡ were evaluated for the first time at this release (Tencent marks them with an asterisk after the name). "n/a" means no score is reported.

Agentic Coding

Benchmark	Hy4 preview	Hy4	GLM-5.3	DeepSeek-V4.1-Flash-1017	Kimi K3.1 Pro	GPT-6.1 Sol	Claude Opus 5.5
DeepSWE v1.1	64.3	74.9	66.9/68.1*	77.0/73.8*	75.5/76.3*	75.2/72.7*	74.2/72.6*
SWE-Marathon v1.1 ‡	31.9	43.8	42.5/35.6*	46.3*	51.7/53.8*	42.5/39.4*	46.3/44.4*
ProgramBench (almost solved) ‡	17.5	30.0	19.0/18.0*	29.1/29.0*	31.5/24.5*	46.0/46.0*	65.0/56.5*
Terminal-Bench 4.0	26.3	44.4	42.9*	42.3/40.0*	65.7/66.8*	55.9/55.5*	66.4/66.4*
FrontierCode 2.0 ‡	n/a	31.0	21.6	28.5	25.5	40.6	46.4
Hy-CompanyBench 2.0 (internal)	62.4	78.8	64.5*	79.5*	70.0*	70.2*	72.3*
Hy-SWE Max Verified 2.0 (internal) ‡	39.2	45.4	37.7*	44.3*	41.3*	45.1*	56.9*
Hy-Backend 3.0 (internal) ‡	23.3	45.1	29.1*	37.3*	39.8*	44.0*	48.4*
WebCraftBench (internal)	52.7	78.9	55.4*	76.1*	83.1*	85.5*	78.8*
GameFix Bench 2.0 (internal)	65.5	68.4	63.3*	67.5*	66.3*	74.0*	74.8*
3DArena-Modeling (internal)	23.4	37.1	23.0*	32.3*	36.1*	41.8*	43.4*
Goal2Repo (internal) ‡	18.1	23.2	15.2*	20.3*	21.0*	58.0*	74.0*
AutoCatBench (internal) ‡	n/a	57.2	51.9*	58.1*	62.6*	65.8*	72.4*

Agentic Search

Benchmark	Hy4 preview	Hy4	GLM-5.3	DeepSeek-V4.1-Flash-1017	Kimi K3.1 Pro	GPT-6.1 Sol	Claude Opus 5.5
Hy-LifeSearch (internal)	45.9	60.2	45.7*	53.0*	52.0*	48.1*	46.2*
Hy-BrowseComp 2.0 (internal)	53.7	59.9	49.3*	55.1*	59.3*	61.7*	60.7*

Working Agent

Benchmark	Hy4 preview	Hy4	GLM-5.3	DeepSeek-V4.1-Flash-1017	Kimi K3.1 Pro	GPT-6.1 Sol	Claude Opus 5.5
APEX-Agents v1.1 (pass@1)	53.9	75.2	64.3*	72.1/72.9*	67.2/68.2*	60.0/61.7*	73.5/74.8*
Agents' Last Exam v1.1 (Pass Rate) ‡	n/a	41.0	26.2	35.3	43.4	40.2	43.4
AutomationBench v1.0.6	36.3	56.2	48.2/54.1*	60.5/54.3*	65.6*	51.4*	58.8*
E-Bench 2.0 (internal)	58.8	70.0	50.3*	66.8*	50.0*	65.3*	80.0*
Agents' First Exam-IF (internal)	50.5	79.9	60.1*	68.7*	79.9*	84.8*	77.9*
Tag-bench (internal)	42.5	48.8	41.5*	48.3*	53.4*	56.7*	58.3*
PPTGen-Bench (internal)	49.0	60.9	53.9*	62.5*	64.5*	66.5*	67.7*

STEM

Benchmark	Hy4 preview	Hy4	GLM-5.3	DeepSeek-V4.1-Flash-1017	Kimi K3.1 Pro	GPT-6.1 Sol	Claude Opus 5.5
HLE-Full (with tools)	57.3†	59.1	62.5/57.4†*	57.1*	70.0/61.8*	60.6*	67.7/66.2*
Hy-FrontierMath T5 (internal) ‡	7.8	13.0	0.0*	8.0*	14.8*	97.5*	91.3*
ProofBench (internal) ‡	3.3	5.5	0.0*	0.3*	3.9*	70.5*	42.2*

AI for Research

Benchmark	Hy4 preview	Hy4	GLM-5.3	DeepSeek-V4.1-Flash-1017	Kimi K3.1 Pro	GPT-6.1 Sol	Claude Opus 5.5
Terminal-Bench-Science 0.1	16.2	19.8	10.0*	13.7*	38.1/41.7*	57.0/60.9*	58.7/56.2*
PostTrainBench v1.2 ‡	31.8	35.2	34.5/31.6*	30.5/29.4*	36.5/36.2*	38.8/38.2*	43.8/44.2*
Foundation Model Bench (internal) ‡	13.6	28.1	25.6*	36.3*	25.1*	41.3*	55.4*
TenureTrail (internal) ‡	46.8	53.6	46.0*	50.1*	51.3*	57.6*	59.0*

Computer Use and Multimodal

Benchmark	Hy4 preview	Hy4	GLM-5.3	DeepSeek-V4.1-Flash-1017	Kimi K3.1 Pro	GPT-6.1 Sol	Claude Opus 5.5
OSWorld 2.1 (Full, partial score)	n/a	78.9	n/a	68.4*	79.6	75.2*	81.8
MMMU-Pro	n/a	78.6	n/a	76.7*	84.1/83.4*	84.9*	88.2*
ZeroBench-main (with tools, pass@5)	n/a	55.4	n/a	51.0	63.0/56.9*	65.3*	63.5*
BuddyVQA Atlas (with tools, internal)	n/a	81.0	n/a	76.2*	80.2*	79.1*	81.3*
BuddyVQA Atlas (without tools, internal)	n/a	55.6	n/a	43.7*	72.2*	72.0*	75.3*

In this table Hy4 has the top score on APEX-Agents v1.1 (75.2) and Hy-LifeSearch, and it scores higher than Hy4 preview on every benchmark where both are reported. Tencent states plainly that Hy4 still has a significant gap to the top closed-source models, especially on its hardest test sets in frontier mathematics, long-horizon development, model iteration and visual question answering.

Benchmark methodology notes (as disclosed by Tencent)

• Certain Hy4 preview scores may differ from previously reported results following updates to the evaluation harness, judge model and anti-hacking mechanisms.
• DeepSWE v1.1. Evaluated with mini-swe-agent. Each task runs in a sandbox with 8 CPU cores and 16 GB of memory.
• SWE-Marathon v1.1. Official settings, except that the agent timeout is twice the official value. To reduce variance each task is run repeatedly until eight valid scores from successful runs are obtained.
• ProgramBench. Tencent reports the Almost Solved score using Claude Code 2.1.220. Other models use the scaffold configurations from their official blog posts or ProgramBench's official mini-SWE-agent when unspecified. Claude Opus 5.5 encountered content filter blocks on 41 instances.
• Terminal-Bench 4.0. Evaluated with the Claude Code harness, allowing up to 500 turns and a 13-hour timeout per trial. Resources are capped at 16 CPUs and 32 GB of RAM, and the three GPU-dependent tasks also receive a single H100 GPU.
• Hy-CompanyBench 2.0, Hy-SWE Max Verified 2.0 and Hy-Backend 3.0. GPT-series models are evaluated in Codex 0.154.0 and other models in Claude Code (2.1.234 for Hy-CompanyBench, 2.1.295 for the other two).
• WebCraftBench. Evaluates agentic text-to-website generation with judge perspectives aligned to human preferences. GPT-series models use Codex 0.115.0 and other models Claude Code 2.1.117. Website artifacts are deployed in an internal sandbox. The exploration agent is built on AWorld and the judge agent is in-house.
• GameFix Bench 2.0. 100 game repair tasks in a 16-core, 32 GB sandbox. Hidden Playwright tests measure the fraction of game issues resolved.
• 3DArena-Modeling. 120 3D modeling tasks ranging from precision to aesthetics in a 16-core, 32 GB sandbox. Elo scores from blind pairwise comparisons are scaled to percentages.
• Goal2Repo. Evaluates long-horizon coding and self-debugging from a self-contained spec. Agents make multiple submissions and get pass-rate feedback with a failure penalty from a hidden test suite until they fully pass or exhaust the budget (avg@3).
• AutoCatBench. Uses web pages to probe model capabilities through crafted challenges spanning 3D graphics and modeling, physics, search, anime, SVG, professional software, maps, lighting effects, water rendering, typography and planning.
• Hy-LifeSearch and Hy-BrowseComp 2.0. All models use Tencent's internal search harness, graded by an LLM judge (per-task rubrics for Hy-LifeSearch, the unique reference answer for Hy-BrowseComp 2.0).
• APEX-Agents v1.1. Evaluated with the official reference agent harness (Loop) and a maximum of 250 steps per trial. Each trial runs the official two-service environment capped at 8 CPUs and 24 GB of RAM and is graded in a separate verifier container capped at 2 CPUs and 8 GB of RAM.
• AutomationBench v1.0.6. Uses the 600 public tasks released by Zapier, evaluated by the official method.
• E-Bench 2.0. Measures whether a model completes day-to-day tasks with MCP tools across multiple business domains. Tasks run in a simple agent loop and scores combine final database-state checks with an LLM judge.
• Agents' First Exam-IF. 101 research tasks with an internal search harness and per-task rubric grading. It tests whether agents follow critical but unstated constraints.
• Tag-bench. Complex, real-world workplace tasks that require understanding context from multiple sources, including messaging, documents, code repositories and model training platforms, scored with an LLM judge and programmatic checks.
• PPTGen-Bench. Evaluates agentic text-to-PPT generation, combining an aesthetic assessment with penalties for critical faults. Artifacts are rendered to PDF with LibreOffice and the agentic judge uses Claude Code 2.1.117.
• Terminal-Bench-Science 0.1. Evaluated with the Claude Code harness, allowing up to 500 turns and a 13-hour timeout per trial, with resources capped at 24 CPUs and 48 GB of RAM.
• PostTrainBench v1.2. Gemini models are evaluated in OpenCode and all other models in Claude Code 2.1.198. Every task runs in a sandbox with an NVIDIA H100 GPU and a 10-hour budget.
• Foundation Model Bench. Measures whether an agent can optimize domain-specific foundation models through training data, architecture and training strategy.
• TenureTrail. Measures whether AI research agents make real progress on 123 executable tasks spanning mathematics, ML, physical sciences, life sciences, Earth and environmental sciences, engineering and social and economic research.
• ZeroBench-main. Evaluated in a network-isolated Jupyter environment that supports image outputs, with a maximum of 80 tool-calling rounds per run. Each problem is run 10 times independently to estimate pass@5.
• BuddyVQA Atlas. Answering questions from images across recognition, counting, comparison, OCR, spatial orientation and localization, focused on cases where answering from the full image tends to fail, such as small text, thin lines and local targets.

Reasoning and Sampling

Like Hy4 preview, Hy4 reasons by default at Tencent's high setting and offers a no-think mode for direct answers. On Run BiOS reasoning is on by default at high and reasoning_effort: "none" switches to direct answers.

On Run BiOS

Hy4 is available on Run BiOS as the serverless model tencent/hy4, with a 1,000,000-token context window, up to 64,000 output tokens, text and image input, function calling, structured outputs and reasoning that can be switched off. Hy4 preview remains available as tencent/hy4-preview.