What Is Tencent Hy4? 770B MoE, 1M Context
Priya Nair
AI Infrastructure Analyst

TLDRTencent Hy4 has 770B total parameters, 49B active, and 1M+ context for coding, office work, game development, and research.
What Is Tencent Hy4? The 1M-Token Open MoE
Tencent Hy4 is Tencent’s flagship Mixture-of-Experts language model for coding, productivity, game development, and scientific research. Officially launched on November 4, 2026, after an initial preview on August 28, 2026, Tencent Hy4 features 770B total parameters, 49B active parameters per token, a context window exceeding 1M tokens, and new native visual understanding and computer-use capabilities. It is available through Tencent applications, Tencent Cloud TokenHub, OpenRouter, and partner gateways, with public model weights scheduled for release.
Key Takeaways
- Tencent Hy4 is a flagship multimodal language model developed by the Tencent Hy Team, officially launched on November 4, 2026.
- The model contains 770B total parameters across 78 layers and activates 49B parameters per token via 256 routed experts and 1 shared expert.
- Its context window spans 1M tokens (up to 1,048,576 tokens) with a 64K maximum output limit.
- Hy4 introduces native visual understanding and computer use (both CLI and GUI interaction) for the first time in the series.
- List API pricing sits at $0.834 per 1M input tokens and $2.501 per 1M output tokens, with gateway discounts and introductory free access windows.
- Benchmarks place Hy4 competitively among leading frontier models, including an 83.18 average score on WorkBuddy, 75.2 on APEX-Agents, and top rank among open models in blind financial evaluations.
What Is Tencent Hy4?
Tencent Hy4 is the fourth-generation flagship model in Tencent’s Hy family. Following an architectural preview released on August 28, 2026 under the name tencent/Hy4-preview, the complete, general-availability version of Hy4 launched on November 4, 2026.
Tencent built Hy4 specifically for complex, real-world productivity tasks rather than positioning it only as a conversational chatbot. The core workloads include software engineering, office document automation, game development, financial analysis, security, and scientific research. The model’s 770B total parameter pool provides extensive memorization and specialized reasoning capacity, while the 49B active parameter count controls inference compute for individual tokens.
Tencent Hy4 is a 770B-parameter MoE model that activates 49B parameters per token and supports a context window exceeding 1M tokens.
The model was initially previewed on August 28, 2026, through Tencent’s Hy4 announcement. With the November 4 launch, Tencent scaled both pre-training and post-training data, added native visual comprehension, and introduced computer-use capabilities trained on sandboxed software environments.
Tencent Hy4 at a Glance
| Specification | Tencent Hy4 |
|---|---|
| Developer | Tencent Hy Team |
| Type | Large language model; Mixture-of-Experts |
| Modality | Multimodal (Text and image input; text output; CLI/GUI computer use) |
| Total parameters | 770B |
| Active parameters | 49B per token |
| Context window | 1M tokens (64K maximum output) |
| Pricing | $0.834 / 1M input, $2.501 / 1M output list ($0.42 / $1.25 promotional and gateway rates) |
| Availability | Tencent Cloud TokenHub, OpenRouter, Vercel AI Gateway, WorkBuddy, CodeBuddy, Yuanbao; open weights coming soon |
| License | Apache License 2.0 (preview weights released; final weights promised soon) |
| Release state | General availability launched on November 4, 2026 (Preview on August 28, 2026) |
The official Hy4 Preview model card identifies the model as a Tencent Hy Team release under Apache 2.0 licensing, detailing deployment configurations for standard inference engines.

Source: @testingcatalog
How Tencent Hy4 Works and What Makes It Different
Sparse Expert Routing
Tencent Hy4’s core architectural foundation is Sparse Expert Routing. The model incorporates 78 backbone layers. The first layer operates as a dense feed-forward network, while the subsequent 77 layers use MoE blocks containing 256 routed experts and 1 shared expert. Each token selects the top 8 routed experts alongside the shared expert.
This structure decouples model capacity from per-token compute. While serving engines must hold the full 770B checkpoint in GPU memory, only 49B parameters fire during any forward pass. The attention mechanism uses Gated DeepSeek Sparse Attention (Gated DSA) paired with IndexCache for cross-layer index reuse, preventing redundant context computation across layers.
1M Context
1M Context defines Hy4's document- and repository-level processing capacity. Spanning up to 1,048,576 tokens alongside a 64K maximum output limit, the context window can ingest massive code repositories, extensive financial disclosures, or hours of operational logs.
Engineering teams must still evaluate retrieval fidelity, tool-call consistency, and latency across deep context lengths. Long contexts carry higher per-call costs, making prompt structure and selective retrieval necessary components of efficient agent design.
Native Multi-Token Prediction
The architecture also incorporates Native Multi-Token Prediction. Documented in vLLM’s implementation notes, the model features a native 10B multi-token prediction (MTP) layer with 0.7B active parameters and draft depth 3.
This built-in speculative decoding layer allows the system to propose and verify multiple tokens per step, significantly boosting serving throughput when integrated into optimized runtimes such as vLLM and SGLang.
Vision and Computer Use
The November 4 release adds native Vision and Computer Use capabilities. Hy4 can parse images, inspect screenshots, and interpret visual interfaces. Coupled with computer-use training across command-line interfaces (CLI) and graphical user interfaces (GUI), Hy4 can interact directly with software environments.
Tencent deployed specialized training sandboxes across hundreds of desktop applications, training the model to control mouse, keyboard, and terminal actions. This includes targeted adaptations for creative and technical software like Blender and CAD.
What You Can Do With Tencent Hy4
Tencent designed Hy4 for multi-step, tool-augmented engineering and operational workflows:
Software engineering and frontend design are central strengths. Hy4 can inspect multi-file repositories, plan refactors, execute code changes, and debug test failures. With visual comprehension, it can analyze site layouts, take screenshots, and implement responsive frontend interfaces directly from reference designs.
Office collaboration and enterprise workflows integrate through Tencent WorkBuddy and WeCom. Hy4 coordinates documents, spreadsheets, presentations, emails, and calendars across ongoing projects. On WorkBuddy benchmarks, it achieved an average performance score of 83.18 and a 96.15% pass rate.
Game development draws on proprietary data from Tencent’s Interactive Entertainment Group (IEG). Hy4 is adapted for major engines including Unreal Engine, Unity, and Godot, handling gameplay logic, state management, and asset pipelines. In evaluations, Hy4 improved by 19.1% on GameDev and 9.8% on GameLogic compared to Hy4 Preview.
Financial analysis leverages domain-specific training on fund research, securities data, and quantitative investment workflows. In a blind evaluation of 579 tasks judged by 213 financial professionals (including 153 external institutional experts), Hy4 achieved top ranking among open models with a 3.04 score, edging out DeepSeek-V4.1-Flash (2.97).
Scientific research applications span mathematics, biology, and physics. During development evaluations, Hy4 completed an automated 3D reconstruction of approximately 6.99 million cells across 84 mouse embryo slices in 1.5 hours without human intervention, and collaborated with university researchers to identify a mathematical conjecture counterexample verified in Lean.
How Tencent Hy4 Compares
Across benchmark evaluations, Tencent Hy4 demonstrates competitive parity with leading open-weight and proprietary models:
| Benchmark / Evaluation | Tencent Hy4 | Comparison Models |
|---|---|---|
| Financial Blind Eval (579 tasks) | 3.04 / 4.00 | DeepSeek-V4.1-Flash: 2.97 / 4.00 |
| APEX-Agents | 75.2 | DeepSeek-V4.1-Flash: 72.1 |
| Spark-Bench (thinking off) | 90.9 | GLM-5.3: 90.3 · DeepSeek-V4.1-Flash: 88.7 |
| OSWorld (GUI computer use) | 78.9 | GPT-6.1 Sol: 75.2 · Claude Opus 5.5: 81.8 |
| Terminal-Bench 4.0 | 44.4 | GPT-6.1 Sol: 55.9 · Claude Opus 5.5: 66.4 |
| Preview Engineering Eval (203 tasks) | 2.99 / 4.00 | GLM-5.3: 2.92 · Kimi K3: 2.94 |
Tencent Hy4 achieved a 3.04 score in expert financial evaluations and 75.2 on APEX-Agents, establishing strong positioning among leading models.
Hy4 shows pronounced strength in agent workflows, game engineering, and GUI-driven software automation, while trailing proprietary frontier models like Claude Opus 5.5 on Terminal-Bench 4.0. In internal hallucination testing, targeted verification optimizations reduced the model's hallucination rate from 15.76% down to 5.45%.
For context on neighboring models in the ecosystem, see our analysis of GLM-5.3 and our coverage of Kimi K3’s release.
Availability: How to Access Tencent Hy4
Tencent Hy4 is available through both API gateways and end-user enterprise applications:
- API Access: Live on Tencent Cloud TokenHub, OpenRouter, and Vercel AI Gateway. Third-party aggregators such as ZenMux and Respan also host API endpoints.
- Application Integration: Available in Tencent WorkBuddy, CodeBuddy, Yuanbao, and ima. WorkBuddy and CodeBuddy provide free access from November 4 to 15, 2026.
- API Pricing: Official list rates are $0.834 per 1M input tokens and $2.501 per 1M output tokens (with cached input at $0.042/M). Vercel AI Gateway and OpenRouter offer discounted pricing at $0.42 input and $1.25 output per 1M tokens, with TokenHub running 50% promotional discounts through December 31, 2026.
- Model Weights: While the August 28 Hy4 Preview weights reside on Hugging Face and Tencent’s Hy4 Preview GitHub repository, Tencent announced that the final Hy4 model weights and safety system cards will be published in the days following the API launch.
Teams requiring large-scale local serving should account for the model's footprint: running unquantized 770B parameters requires multi-GPU clusters (such as 16x B200 or 8x B300 setups), though FP8 and advanced quantization schemes (like Tencent's Sherry method) compress requirements considerably.
For teams exploring alternatives while infrastructure configurations are established, Kimi K3 represents an established alternative. Kie.ai does not currently host Tencent Hy4.
What We Don’t Know Yet
A few specific technical details remain open following launch:
- The exact release date for downloadable weights and system cards for the final multimodal checkpoint.
- Independent verification of the reported hallucination reduction metrics (15.76% to 5.45%) across external academic benchmarks.
- Broad independent community replication of extended multi-turn computer-use latency and reliability on complex desktop environments.
- Detailed architecture specifications and release dates for the announced lightweight companion model, Hy4 Flash.
Frequently Asked Questions
What is Tencent Hy4?
Tencent Hy4 is Tencent's flagship Mixture-of-Experts AI model for coding, productivity, game development, office workflows, and scientific research. Officially launched on November 4, 2026 following an August 2026 preview, it features 770B total parameters, 49B active parameters per token, a 1M-token context window, and native vision and computer-use capabilities.
Is Tencent Hy4 open source?
The August 2026 Hy4 Preview was released with open weights under the Apache License 2.0. The full Hy4 model launched on November 4, 2026 via official cloud APIs and enterprise applications, with Tencent committing to release the official model weights and system cards in the coming days.
How many parameters does Tencent Hy4 have?
Tencent Hy4 has 770B total parameters and activates 49B parameters per token. Its Mixture-of-Experts backbone spans 78 layers with 256 routed experts and 1 shared expert, alongside a 10B native multi-token prediction layer.
What is Tencent Hy4's context window?
Tencent Hy4 has a context window of 1M tokens (up to 1,048,576 tokens) with a maximum output limit of 64K tokens, designed for full repositories, multimodal inputs, and extended agent sessions.
How much does Tencent Hy4 cost?
Tencent Hy4 list API pricing is $0.834 per million input tokens and $2.501 per million output tokens. Gateways like Vercel AI Gateway and OpenRouter offer discounted access at $0.42 input and $1.25 output per million tokens, Tencent TokenHub runs a half-price promotion through December 31, 2026, and access on WorkBuddy and CodeBuddy is free from November 4 to 15, 2026.
How can I access Tencent Hy4?
Tencent Hy4 is accessible via API on Tencent Cloud TokenHub, OpenRouter, and Vercel AI Gateway, and directly in apps such as WorkBuddy, CodeBuddy, Yuanbao, and ima. The final open weights are scheduled for public release shortly after launch.
Tencent Hy4 vs GLM-5.3: which is better?
Tencent Hy4 scored slightly above GLM-5.3 in Tencent's initial blind engineering evaluation (2.99 vs. 2.92 out of 4.00) and achieved 90.9 on Spark-Bench compared to GLM-5.3's 90.3. However, performance varies depending on workload and tooling.
What to Watch Next
Key developments to track include the public release of the final Hy4 weights, the launch of the faster Hy4 Flash model, and broader community evals of its GUI computer-use agents.
Try related models on Kie.ai: DeepSeek-V4.1-Flash, Kimi K3 and Claude Opus 5.5.
About Priya Nair
Priya covers serving costs, context windows, and the infrastructure tradeoffs behind each model launch.
View all posts by Priya Nair