HY4 is Tencent’s new flagship AI model, released on November 4, 2026, for multimodal reasoning, coding, computer use, research, and professional workflows. It adds vision and GUI interaction to the Hy model family and improves sustained execution across documents, software, and business tools. Tencent reports scores of 74.9 on DeepSWE v1.1, 75.2 on APEX-Agents v1.1, and 78.9 on OSWorld 2.1 (full, partial score), under its published evaluation conditions.
Key Takeaways
Strong real-world gains vs preview: +21 points on APEX-Agents, major lifts on Terminal-Bench 4.0, AutomationBench, finance expert scores, and product-level metrics.
Independent ranking: #4 open-weight on Vals Index (59.35%), #2 open-weight on Terminal-Bench 4.0 in some reports, competitive on Finance/Tax Agent.
Reviewed for clarity, source attribution and current API terminology.
Vision + computer use added for the first time, with specialized training on professional software (Blender, CAD, etc.).
Pricing remains highly competitive at ~$0.83/$2.50; CometAPI often offers further discounts (~20%).
RSI signals: Hy4 + Hyra 2.0 already used to improve internal training pipelines and co-develop open models (HyCell, HyPDE).
Worth it for: long-horizon agents, finance/office workflows, multimodal productivity, and cost-sensitive teams that need strong open-model performance.
Not the absolute leader on every hard STEM or pure coding leaderboard — still a clear step below the very top closed models on those.
What Is HY4, and Why Does It Matter?
HY4, styled Hy4 by Tencent, is a general-purpose AI model designed to carry out complex work over many steps. Tencent’s November 4 announcement positions it around a combination of reasoning, visual understanding, tools, and sustained execution.
That combination matters because professional tasks rarely fit into one isolated question. A developer may need an agent to inspect a repository, implement a feature, run tests, and revise a UI. An analyst may need it to reconcile filings, spreadsheets, charts, and research notes. A workplace assistant may need to connect messages, documents, calendars, and meeting records.
HY4’s release focuses on those workflows. Tencent says it substantially increased both pre-training and post-training relative to its earlier version, with gains in instruction following, long-context understanding, deep research, and long-horizon tasks.
Core architecture (inherited and extended from preview)
Sparse Mixture-of-Experts: large total parameter count with only ~49B activated per token.
1M-token context window for repository-scale and multi-document work.
Gated DeepSeek Sparse Attention + IndexCache for efficient long-context serving.
Identity Hyper-Connections and native multi-token prediction (MTP) layer for speculative decoding.
New multimodal pathway and computer-use action space (screen understanding + CLI/GUI control), with targeted training on professional tools and cross-application workflows.
Recursive Self-Improvement (RSI) signals
Tencent reports that Hy4, paired with the Hyra 2.0 harness, optimized internal pre-training infrastructure for smaller same-architecture models and autonomously produced variants that outperformed official baselines. Two foundation models co-developed this way (HyCell for single-cell representation and HyPDE for physical-field modeling) are being open-sourced. This is one of the more concrete public examples of a frontier model contributing to its own development loop.
What performance in the Full HY4 Release?
Vision and Computer Use
HY4 introduces visual perception, including fine-grained image understanding and complex-chart interpretation. Tencent also adds computer use across command-line interfaces and graphical interfaces, with targeted training for software such as Blender and CAD. Tencent’s release describes workflows spanning multiple applications.
These are related but distinct capabilities. Vision lets a model interpret a screenshot or chart. Computer use adds a feedback loop: inspect the interface, choose an action, observe the result, and adjust. It requires a surrounding application that supplies screenshots and executes actions; a text completion endpoint alone does not provide an operating-system agent.
Frontend Engineering and Game Development
Tencent emphasizes stronger UI interactions, business logic, state management, and full-stack integration. For games, it worked with developers from Tencent Interactive Entertainment Group on feature implementation, gameplay logic, state tracking, and asset pipelines.
The release reports 19.1% improvement on GameDev and 9.8% on GameLogic over the earlier version, and describes adaptation to Unreal Engine, Unity, and Godot. These are Tencent’s reported improvements on named evaluations. They do not establish an equivalent reduction in development time for every team.
For developers, the relevant test is whether an agent delivers a working feature with fewer corrections: rendered appearance, correct interactions, persistent state, executable code, and passing tests.
Finance, Office Work, and Personal Assistance
Tencent built thousands of realistic financial workspaces and tens of thousands of tasks covering investment research, quantitative strategies, allocation, corporate finance, and tax accounting. Its evaluation involved 213 finance professionals, including 153 experts from leading financial institutions, across 579 tasks; More than 70% of the experts had over five years of experience.
HY4 scored 3.04, compared with 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. These results come from Tencent’s evaluation, rather than a separate independent replication.
Tencent also describes real-time co-editing of Word, Excel, and PowerPoint files. Through WorkBuddy and WeCom, HY4 can use connected workplace context to prepare briefs, track project changes, identify action items, and flag deadlines. Those integrations belong to the product environment; they are not automatically included in every HY4 API call.
HY4 Performance Benchmarks: What Does the Data Show?
The following comparison uses the full HY4 release’s benchmark appendix. All entries are Tencent-published; where competitors have two scores, the table uses Tencent’s own retest, marked with an asterisk in the original. It does not combine the best result from different runs.
Benchmark
HY4
GLM-5.3
DeepSeek-V4.1-Flash-1017
Kimi K3.1 Pro
GPT-6.1 Sol
Claude Opus 5.5
DeepSWE v1.1
74.9
68.1
73.8
76.3
72.7
72.6
Terminal-Bench 4.0
44.4
42.9
40.0
66.8
55.5
66.4
APEX-Agents v1.1, pass@1
75.2
64.3
72.9
68.2
61.7
74.8
AutomationBench v1.0.6
56.2
54.1
54.3
65.6
51.4
58.8
OSWorld 2.1, full, partial score
78.9
—
68.4
79.6
75.2
81.8
MMMU-Pro
78.6
—
76.7
83.4
84.9
88.2
TenureTrail, internal
53.6
46.0
50.1
51.3
57.6
59.0
Independent evaluations (as of early November 2026)
Vals Index: 59.35% (#13 overall, #4 open-weight behind Kimi K3.1 Pro, DeepSeek V4.1 Flash, GLM-5.3). Big improvement from Hy4 preview’s ~53%. Strong on Terminal-Bench 4.0 (#2 open-weight in some reports), Finance Agent, Tax Agent.
Artificial Analysis Intelligence Index: ~34.6–35 (solid mid-tier, verbose, slower than average per task).
Plotly Data Analytics Bench: 42/43 (98%), among the cheapest models to reach that accuracy level.
Agent Arena community data: strong real-user long-horizon performance; Hy4 quickly became the #1 Chinese model and sits on the Pareto frontier at competitive cost per task.
Where HY4 Looks Strong
HY4’s 75.2 on APEX-Agents v1.1 is the highest entry among the displayed comparisons. Its 74.9 on DeepSWE v1.1 is also competitive with the other models in Tencent’s retests. For scientific workflows, 53.6 on TenureTrail leads the open-model competitors shown, although the two proprietary models score higher.
These findings support testing HY4 for business agents, repository work, and research. They do not support a universal “best AI model” claim.
Where Competitors Still Lead
HY4’s 44.4 on Terminal-Bench 4.0 trails the shown Kimi and Claude results. Its 78.6 on MMMU-Pro also trails Kimi, GPT, and Claude in the table. Computer use is competitive, with 78.9 on OSWorld 2.1, but the metric is explicitly a partial score, not a claim that 78.9% of arbitrary desktop jobs will complete successfully.
Tencent acknowledges a remaining gap with leading proprietary systems on research mathematics, long-horizon coding, model self-improvement, and advanced visual question answering.
Why Evaluation Settings Matter
Tencent’s Terminal-Bench 4.0 setup allows up to 500 turns and a 13-hour timeout, with 16 CPUs and 32 GB RAM; GPU-dependent tasks additionally receive an H100. Its APEX-Agents setup allows up to 250 steps. These are substantial agent budgets.
A model evaluated inside that environment may behave differently in a web app with a 30-second timeout. Compare task success alongside total latency, tool costs, token use, and human correction. Keep benchmark versions visible: Terminal-Bench 4.0 scores should not be read as directly interchangeable with older Terminal-Bench results.
How HY4 Works: Context, Reasoning, and Agent Execution
Long Context Helps Connect Distributed Information
HY4 with a 1,048,576-token context window, text-and-image input, and text output. It lists a maximum completion of 65,536 tokens for its top-provider configuration. These are endpoint-specific limits; confirm CometAPI’s current limits before designing around them.
Long context can help an agent connect information across files and tool traces. It does not guarantee perfect recall, accurate arithmetic, or correct source selection. A well-designed application still benefits from retrieval, document identifiers, concise tool results, and explicit citations.
Reasoning and Tool Use Form a Feedback Loop
A typical HY4 workflow combines planning with external evidence: read the task, request information, inspect a tool result, run a calculation or test, and revise the output. The model’s reasoning helps select the next step; the execution environment supplies current facts and checks.
For example, a financial assistant can retrieve statements, build a valuation workbook, check formulas, and generate a report. Each stage should preserve source references and distinguish reported values from assumptions. A coding assistant can inspect files, change code, run tests, and use failures to guide the next edit.
Architecture Details Require Careful Versioning
The widely quoted 770B total parameters and 49B active parameters describe HY4 preview’s published MoE backbone. Its model card also documents sparse attention, IndexCache, identity Hyper-Connections, and multi-token prediction.
Tencent’s full-release announcement says additional technical details will accompany the weights and system card. Until those arrive, the preview’s layer counts, expert configuration, and deployment commands should not be presented as confirmed full-HY4 specifications. The established full-release facts are its new capabilities, hosted access, and published evaluation results.
HY4 and RSI: What Recursive Self-Improvement Actually Means
From Solving a Problem to Improving the Tool
RSI stands for recursive self-improvement. In Tencent’s November 2 Hyra 2.0 announcement, the idea is to improve reusable research tools, training pipelines, and models. Stronger tools can help develop stronger models, which can support the next round of tool development.
The full HY4 release gives a concrete example: paired with Hyra 2.0, HY4 optimized internal pre-training infrastructure for smaller models with the same architecture and trained a variant that surpassed the stated Jev baseline on multiple benchmarks. Tencent also announced HyCell and HyPDE, foundation models for single-cell representation and physical-field modeling co-developed with the system. Source.
How Hyra 2.0 Organizes Research
Hyra uses three levels: Overall, Mission, and Worker. Overall coordinates human goals, priorities, and resources. Missions direct longer research efforts. Workers inspect code, run experiments, analyze failures, and challenge hypotheses.
A persistent task queue handles dependencies and resource limits. A Research Board records evidence and decisions. Review Workers independently inspect results, while shared knowledge separates observations from reproduced findings and team conventions. Tencent says substantial compute spending and shared-knowledge changes require appropriate human approval.
What RSI Does—and Does Not—Establish
This is evidence of a structured AI research workflow. It does not establish unlimited intelligence growth or show that a hosted HY4 conversation changes the endpoint’s underlying weights. Improved prompts, persistent memory, new software tools, and a newly trained model are different forms of improvement.
For developers, the useful pattern is measurable iteration: preserve failures, improve tools, evaluate candidates on held-out tasks, and promote changes only when the evidence supports them. HY4’s 35.2 on PostTrainBench v1.2 and 28.1 onTencent’s internal Foundation Model Bench are relevant indicators, but remain evaluation-specific.
Reliability: Has HY4 Reduced Hallucinations?
Tencent reports that targeted search and verification improvements reduced its internal hallucination rate from 15.76% to 5.45%. That is a 10.31-percentage-point decrease, or approximately 65.4% relative reduction, calculated from Tencent’s figures. Speculative errors fell from 23.1% to 5.6%. Source.
Those results concern Tencent’s finance and office evaluations. They should not be generalized into a claim that HY4 hallucinates only 5.45% of the time across all questions.
There is supporting product feedback: RabyteTech reports 61.63 on 101 investment-research tasks, up 12.03 points over the earlier version. Manus reports nearly 97% completion on its internal document creation and editing tasks. These are partner testimonials reproduced in Tencent’s announcement, with different task sets and definitions; they are useful context rather than interchangeable benchmark scores.
HY4 Pricing: Official Rates, CometAPI, and Provider Offers
Current Price Comparison
The following snapshot was checked on November 4, 2026. All rates are USD per one million tokens.
Access route
Uncached input
Output
Cached input
Tencent release-page reference
$0.834
$2.501
$0.042
CometAPI hy4 listing
$0.6672
About $2.00
Not verified on the model page
OpenRouter tencent/hy4 API metadata
$0.417
$1.2505
$0.021
CometAPI displays a 20% discount against its listed official reference price. OpenRouter’s live API metadata shows a lower rate than Tencent’s release-page reference at the time checked. The metadata does not establish how long that offer lasts. These differences make live provider-level verification essential; CometAPI should not be described as the lowest-priced route on this evidence.
Estimate the Cost of Your Actual Workload
At CometAPI’s displayed rates, a workload using 1M input tokens and 250,000 output tokens costs approximately:
$0.6672 + (0.25 × $2.00) = $1.1672
This is an illustrative token calculation. Retries, reasoning-token accounting, caching, tool services, and account-specific terms can affect the final bill. Output-heavy agents can spend more than expected when they repeatedly inspect files or verify work.
Compare cost per successfully completed task. A lower token rate may lose its advantage if a model needs more attempts, longer outputs, or substantial human repair.
Conclusion
Hy4 delivers meaningful, measurable performance gains over its preview while preserving the aggressive pricing that made the earlier release attractive. Independent rankings confirm it sits in the top tier of open-weight models, with particular strength in agents, finance, and real productivity tasks. Combined with new vision and computer-use abilities, it is currently one of the most compelling options for cost-conscious teams that need strong open-model capability.
For the simplest, most flexible way to evaluate and deploy it alongside other leading models, route through CometAPI. One key, competitive pricing, and seamless switching make it the practical choice for production and experimentation in 2026.
FAQs About HY4
What Is Tencent HY4?
HY4 is Tencent’s flagship AI model released on November 4, 2026. It supports visual understanding, computer use, reasoning, coding, research, and multi-step professional tasks.
What Changed from the Earlier Version?
The full HY4 adds vision and computer use and scales pre-training and post-training. Tencent reports improvements in long-context understanding, instruction following, finance, office work, development, and scientific research.
What Are HY4’s Main Benchmark Scores?
Tencent reports 74.9 on DeepSWE v1.1, 44.4 on Terminal-Bench 4.0, 75.2 on APEX-Agents v1.1, 78.9 on OSWorld 2.1 partial score, and 78.6 on MMMU-Pro. These use the release appendix’s evaluation conditions.
Does HY4 Have a 1M-Token Context Window?
OpenRouter’s live metadata lists 1,048,576 tokens for HY4. Check the limits of the endpoint you use, including the combined input/output budget and maximum completion length.
Is HY4 Self-Improving?
Tencent is developing RSI workflows with HY4 and Hyra 2.0 to improve research tools, infrastructure, and model variants. A normal hosted conversation does not demonstrate autonomous modification of HY4’s deployed weights.
How Much Does HY4 Cost on CometAPI?
The checked model page lists $0.6672 per 1M input tokens and approximately $2.00 per 1M output tokens, with a displayed 20% discount against its official reference. Consult the live price page before budgeting.
Is HY4 Open Source and Ready to Download?
Tencent says full HY4 weights and a system card are planned. Public preview weights are a separate release. Confirm the full model’s actual download availability and license before self-hosting.
Can I Use HY4 Through CometAPI Now?
Yes. CometAPI lists hy4 as available. Use its API documentation to confirm account access, parameters, image support, and rate limits, then run a small workload before scaling.
