Tencent Hunyuan 3: The New AI Model Powerhouse

Abstract

Tencent Hunyuan 3 (Hy3) official version has been fully launched after half a year of intensive R&D investment by Tencent’s AI team led by Yao Xiaoyu, the newly appointed Chief AI Scientist. This paper systematically sorts out the core R&D background, architectural parameters, standardized benchmark comparison data, real industrial scenario test indicators, pricing & open-source policies, and large-scale commercial landing results of Hy3. Comprehensive horizontal comparisons are conducted with mainstream open-source and closed-source LLMs including GLM-5.2, DeepSeek V4 Pro, GPT-5.5 and Gemini 3.1 series. Enterprises operating multi-model business workloads can uniformly schedule model traffic via an API gateway such as 4sapi to simplify hybrid model access management.

1. R&D Background & Core Reform of Tencent’s Large Model Infrastructure

In December 2025, Tencent completed a comprehensive overhaul of its internal large-scale model R&D architecture, establishing three core foundational platforms: AI Infra, AI Data and data computing infrastructure. The group recruited Yao Xiaoyu, a top scholar formerly from Tsinghua University, to serve as Chief AI Scientist, reporting directly to Liu Pinghan and Lu Shan.

Yao’s first core adjustment after taking office was to discard the original training framework. Within one month, the team rebuilt the full set of training and reinforcement learning infrastructure, and established three core development principles: no bias tuning, no brute-force overfitting, and no waste of computing resources.

The first deliverable after the architectural reconstruction was Hy3 Preview, launched on April 23. The preview version only took three months from cold-start training to release, yet it was merely a trial product with limited comprehensive capabilities, failing to reach domestic SOTA level. The newly released official Hy3 version completes full capability iteration and truly achieves industry-leading comprehensive performance.

2. Core Architecture & Parameter Specifications of Official Hunyuan 3

The official Hy3 inherits the underlying framework of its preview iteration, with fixed core hyperparameters as follows:

• Total parameter volume: 295B
• Single inference activation parameters: 21B
• MTP (Multi-token Prediction) auxiliary parameters: 3.8B
• Network layers: 80 layers (excluding MLP auxiliary layers)
• GQA grouped query attention mechanism, 64 attention heads including 8 key-value heads
• Hidden dimension: 4096, intermediate hidden dimension: 13312
• Expert mixture configuration: 192 experts, top-8 activation per token
• Max context window: 256K tokens
• Vocabulary size: 120,832, BF16 precision storage

The core structural framework has no changes compared with Hy3 Preview, consistent with GLM-5.2’s scale. The core optimization of the official version lies in richer training data diversity and expanded RL (Reinforcement Learning) scale. Simply put, the architecture remains unchanged, while the training corpus quality, data volume and reinforcement learning computing resource allocation have been comprehensively upgraded.

3. Full Benchmark Horizontal Comparison Data

Tencent released complete standardized test results on its official blog and Hugging Face, covering six major tracks: code generation, information retrieval, agent workflow, STEM science, logical reasoning and long context learning. All test data adopts unified horizontal evaluation standards for mainstream models.

3.1 Code Generation Benchmarks

Hy3 Official achieves 78.0 points on SWE-Bench Verified, 57.9 on standard SWE-Bench Pro, and 75.8 on multilingual SWE-Bench Multilingual; Terminal-Bench 2.1 reaches 71.7 points, DeepSWE 28.0. Horizontal reference: GPT-5.5 scores 84.4 on SWE-Bench Verified, 58.6 on SWE-Bench Pro; GLM-5.2 hits 62.1 on SWE-Bench Pro, DeepSeek V4 Pro scores 55.4. Hy3 is comparable to mainstream open-source models, with a slight gap against top closed-source flagship models.

3.2 Search & Retrieval Intelligence

This is Hy3’s strongest capability track:

• BrowseComp: 84.2, ranking first among all tested models, on par with GPT-5.5
• WideSearch: 76.4
• DeepSeekQA: 91.0

3.3 Agent Workflow Capability

• MCP Atlas (open public set): 79.1
• ClawEval (pass³): 68.5
• Toolathlon: 48.5
• WildClawBench (35-round pure text): 53.6
• Internal financial modeling benchmark Hy-FinModelBench: 69.0, roughly on par with GLM-5.2

3.4 STEM & Mathematical Reasoning

• GPQA Diamond: 90.4 (GPT-5.5: 93.6)
• HLE tool-aided pure text: 53.2 (GLM-5.2: 54.7)
• DeepSeek V4 Pro: 48.2
• USAMO 2026 mathematical competition: 72.0
• IMOAnswerBench Olympiad math: 90.0
• MathArena advanced math: 38.7
• SuperChem chemistry benchmark: 54.9

Hy3’s 90.4 score on GPQA Diamond is close to GPT-5.5’s 93.6; its HLE score is slightly lower than GLM-5.2 but higher than DeepSeek V4 Pro.

3.5 Long Context Learning (Domestic Self-built CL-Bench & AA-LCR)

Two self-developed evaluation suites measure long-context knowledge extraction and reasoning, a core pain point Yao Xiaoyu’s team prioritized optimizing:

• CL-bench: 23.8
• CL-bench Life long text scenario: 17.0
• AA-LCR cross-document reasoning: 73.4

The paper points out that most mainstream LLMs perform poorly on long context tasks; the top-performing GPT-5.5 (High) only reaches 23.7 on CL-bench, while most models cannot solve more than 1% of test tasks. Before Hy3, Claude Opus 4.8 scored only 24.8 on this benchmark, making Hy3’s 23.8 the top result among domestic large models.

4. Real-World Industrial Scenario Stress Test Results

Tencent acknowledges that public benchmarks cannot fully reflect real production performance. Before official release, the team conducted blind real-scenario stress tests with 270 internal business teams, collecting 312 valid double-blind comparison samples covering front-end development, data storage, CI/CD and other fields. The test logic adopts real business expert scoring standards, which better reflect practical production performance than automated benchmarks.

Core improvement indicators compared with Hy3 Preview:

1. Hallucination rate dropped from 12.5% to 5.4%, a reduction of over half
2. Common-sense error rate decreased from 25.4% to 12.7%
3. Multi-round question deviation rate down from 17.4% to 7.9%
4. Long dialogue reasoning accuracy MCR2 increased from 42.9% to 75.1%
5. Tool call error recovery efficiency significantly improved, invalid loop trigger calls reduced
6. Cross-framework generalization enhanced: consistent stable effects when invoking Hy3 via any coding tool framework

5. Pricing Strategy & Open-Source License Rules

5.1 Official API Billing Standard

Hy3 maintains the cost-effective pricing route established by the preview version:

• Input token: $4 per 1 million tokens
• Output token: $4 per 1 million tokens
• In-cache input token: $0.25 per 1 million tokens

5.2 Open-Source Scope

The model weight adopts Apache 2.0 license, fully open to global developers on GitHub, Hugging Face, ModelScope, GitCode and other platforms for free commercial use. Overseas distribution channels including OpenRouter, Cline, OpenClaw, OpenCode and CherryStudio will launch access sequentially.

5.3 Traffic Growth Data

During the preview iteration, Hy3 token calls reached 10 times the volume of the previous generation model, ranking top two in total traffic and market share on OpenRouter. After the official version launched, total token consumption increased by 20 times year-on-year, with explosive growth in code and Agent scenarios: WorkBuddy/CodeBuddy and QLCla class tools recorded over 16.5x traffic growth.

6. Large-Scale Commercial Landing & Vertical Product Performance

Model benchmark scores are only valuable when implemented into end products. Upon official release, Hy3 has been integrated into Tencent’s core business lines including WorkBuddy/CodeBuddy, Yuanbao, ima, Marvis, QQ Browser, Tencent News, WeGame, Tencent Music, Sogou Input, WeChat Official Accounts, WeChat Reader and Tencent Maps, with nearly 50 additional businesses queuing for access.

6.1 WorkBuddy (Office & Coding Agent)

The core verification scenario for Hy3’s capability iteration:

• Task completion rate rose from 72% to 90% vs Preview
• Average response latency reduced by 34%
• Token consumption for high-frequency office tasks dropped significantly vs GLM5.2: document analysis -47.4%, PPT production -49.0%
• Active user volume increased 6x since preview launch; the showcase can generate Excel multi-dimensional analysis, 30-page PPTs, cross-regional data aggregation tables with over 5,000 single cells, energy concept design, image carousel tourist mini-programs and interactive casual games end-to-end.

6.2 Yuanbao (General AI Assistant)

Supports full Agent workflow: users submit natural language demands, and Yuanbao independently completes complex tasks to generate PPT, Word, Excel, PDF, HTML deliverables. Internal evaluation shows Hy3 surpasses GLM-5.1 in comprehensive office and life scenario indicators, with document synthesis accuracy up 7%.

6.3 ima (Knowledge Base & Multi-Agent Collaboration)

• Agent system stability: 95.1%
• Tool writing capability greatly enhanced, blind retries and invalid termination operations cut drastically
• Knowledge base QA hit rate +19%, hallucination rate -15 percentage points
• Marvis multi-agent collaboration task completion rate: 93.7%, 6-agent collaborative task correct execution rate 92%

6.4 Consumer End Products

• QQ Browser code generation task throughput +37.6%
• WeChat Official Account AI customer service intent recognition accuracy improved from 98.28% to 98.94%
• WeGame game assistant multi-round reasoning & tool scheduling comprehensive success rate 92%, hallucination rate down from 4.5% to 2.8%

6.5 WeChat Mini Program Native Development Capability

Hy3’s unique strength lies in end-to-end mini-program engineering. The preview version already supported full-stack mini-program scaffolding from natural language prompts: complete route logic, classification navigation, detail pages, image libraries and local storage, outputting full app.json and executable project files in one generation. The official version further strengthens this capability, supporting full front-end, back-end, API, data structure and project schema generation in one batch.

For reference, Tencent’s previous mini-program assistant "Xiaowei" adopted WLM + DeepSeek hybrid scheduling, while Hy3 realizes full native standalone mini-program development without third-party model assistance. Tencent has reached AI Agent cooperation with Meituan, Didi, JD and other major platforms, and recommends developers use Hy3 for mini-program development, with WLM + DeepSeek as an alternative lightweight combination for ordinary users.

7. Conclusion

The official release of Tencent Hunyuan 3 marks a comprehensive upgrade of Tencent’s self-developed large model stack in long context reasoning, tool invocation, multi-agent collaboration and code generation. Built on a rebuilt training infrastructure and expanded reinforcement learning scale, Hy3 achieves competitive scores across mainstream academic benchmarks, with more prominent advantages in real industrial blind testing scenarios, especially office Agent, code development and search retrieval workloads.

The open Apache 2.0 license and cost-effective tiered API pricing lower the threshold for global developers to adopt the model, while widespread integration into Tencent’s internal consumer and enterprise business lines verifies its stable production-grade capability. For enterprise platforms running mixed multi-model Agent fleets, unified traffic scheduling via platforms like 4sapi simplifies cross-model credential management and consumption statistics for large-scale Hy3 deployment across distributed teams.

Hy3’s core competitive advantage lies in balanced comprehensive performance, domestic long-context optimization and end-to-end engineering workflow support such as mini-program full-stack generation, filling the gap between lightweight open-source models and overseas flagship closed-source LLMs for Chinese enterprise production scenarios.