Hunyuan

Hunyuan is the family of large language models developed by Tencent Hunyuan, the AI division of the Chinese internet conglomerate Tencent, with a text LLM line that includes the open-weights Hunyuan-Large (389 billion parameters, 52 billion active), the Hunyuan-T1 reasoning model, the Hunyuan 2.0 flagship, and the April 2026 Hy3 open-weights release, alongside separate model families for image generation (Hunyuan-DiT, HunyuanImage 3.0), video (HunyuanVideo), and 3D-asset creation (Hunyuan3D). The models are distributed through Tencent Cloud's API, the Tencent Yuanbao consumer assistant, WeChat integrations, and open-weights releases on Hugging Face. As of May 2026, the Hy3 preview (295 billion total parameters, 21 billion active) is Tencent's most capable open-weights text LLM, positioned as a cost-efficient alternative to DeepSeek-V3 in the Chinese and international open-weights market.

At a glance

• Lab: Tencent Hunyuan
• Released: Hunyuan-Large: November 2024; Hunyuan-T1: March 2025; Hunyuan 2.0: December 2025; Hy3 preview: April 23, 2026
• Modality: Text (primary); image, video, and 3D covered by parallel model families
• Open weights: Partial. Hunyuan-Large, Hunyuan-A13B, and Hy3 preview are open-weights on Hugging Face. Hunyuan 2.0 (Think and Instruct) and Hunyuan-T1 are available via Tencent Cloud API and the Yuanbao consumer product, not as downloadable weights.
• Context window: 256K tokens (Hunyuan-Large, Hunyuan 2.0, Hy3 preview); 128K for the Hunyuan-Large Instruct variant
• Pricing: Tencent Cloud API for HY2.0 Instruct: approximately RMB 0.0045 per thousand input tokens and RMB 0.011 per thousand output tokens (post-March 2026 increase). Hy3 preview weights are free for self-hosting. Hunyuan-A13B via API at approximately $0.14 per million input tokens. First-time Tencent Cloud users receive 1 million free tokens on Hunyuan-turbo.
• Distribution channels: Tencent Cloud API at cloud.tencent.com; Tencent Yuanbao consumer assistant; WeChat embedded AI features; Hugging Face tencent organization for open-weights releases; GitHub Tencent-Hunyuan for model code and documentation

Origins

Tencent's Hunyuan program launched in 2023 in response to the generative-AI commercial wave. The first publicly announced Hunyuan large language model appeared in mid-2023, followed by multimodal vision extensions and an image-generation line through 2024.

The November 2024 Hunyuan-Large release was the first significant open-weights milestone. Hunyuan-Large (also called Hunyuan-MoE-A52B) shipped 389 billion total parameters with 52 billion active, trained on seven trillion tokens including 1.5 trillion tokens of synthetic data emphasizing mathematics, coding, and multilingual coverage. It claimed to outperform Llama 3.1-405B on MMLU and MATH benchmarks while running at lower active-parameter cost. The architecture used Grouped Query Attention and Cross-Layer Attention to reduce key-value cache memory overhead, and supported sequences up to 256K tokens in the pretrain variant and 128K in the Instruct variant.

The March 2025 Hunyuan-T1 release introduced the hybrid Mamba-Transformer mixture-of-experts architecture, built on the TurboS fast-thinking base. The hybrid design was intended to deliver roughly twice the decoding speed of a standard Transformer at equivalent parameter count. T1's published MMLU-Pro score of 87.2 placed it at rank 2 at release, behind OpenAI's o1 and comparable to or slightly above DeepSeek-R1. Tencent allocated 96.7% of T1's computing budget to reinforcement learning rather than supervised fine-tuning.

Hunyuan 2.0, released December 2025, expanded the text flagship to 406 billion total parameters and 32 billion active parameters, with a 256K context window and separate Think (reasoning) and Instruct (chat) variants. The release was positioned as the commercial API flagship and as the backbone of the Yuanbao consumer product.

In February 2026, Tencent rebuilt the Hunyuan infrastructure from scratch. The result, Hy3 preview, was released open-weights on April 23, 2026, approximately 90 days later. Hy3 is a 295-billion-parameter MoE with 21 billion active parameters and a 256K context window. The model uses a dense-MoE hybrid architecture with 192 routed experts and one always-active shared expert per MoE layer, routing tokens of varying difficulty to experts with different capacities.

Capabilities

The Hunyuan text LLM line covers instruction-following, multi-turn dialogue, code generation, mathematical reasoning, and agent tasks. The Hy3 preview, as the most recent release, is Tencent's current open-weights priority.

Hy3 introduces a fusion reasoning mode that integrates fast and slow thinking into a single model rather than maintaining separate chat and reasoning releases, following the direction taken by OpenAI's o-series. The model prioritizes STEM benchmarks and complex agent tasks including web research and software engineering.

Hunyuan-A13B is an 80-billion-parameter total MoE with 13 billion active parameters, available open-weights on Hugging Face in pretrain, instruct, FP8, and GPTQ-Int4 variants, positioned as a cost-efficient alternative for teams that cannot run the full Hy3 at scale.

The broader Hunyuan family covers three modalities beyond text. Hunyuan-DiT and HunyuanImage 3.0 cover text-to-image generation, with HunyuanImage 3.0 the largest open-source image generation MoE by parameter count (80 billion total, 13 billion active). HunyuanVideo (13 billion parameters, December 2024) covers video generation, with extensions for image-to-video, avatar animation, and customized video following in 2025. Hunyuan3D covers text-to-3D and image-to-3D asset generation through Hunyuan3D-2 and 2.5. This profile focuses on the text LLM line.

Benchmarks and standing

Benchmark data for the Hunyuan text line is spread across multiple releases, and third-party coverage is thinner than for DeepSeek or Alibaba Qwen.

For Hunyuan-T1 (March 2025): MMLU-Pro 87.2 (rank 2 at release, behind o1); GPQA Diamond 69.3; MATH-500 96.2; LiveCodeBench 64.9; ArenaHard 91.9. Tencent characterized T1 as outperforming GPT-4.5 and DeepSeek-R1 on MMLU-Pro at release, though independent corroboration of those claims was less comprehensive than coverage of DeepSeek flagship releases.

For Hunyuan-TurboS (the fast-thinking base underlying T1, benchmarked separately): MMLU 89.5; MMLU-Redux 87.11; TriviaQA 92.22; SuperGPQA 54.63 (reported by Tencent as first among contemporaneous competitors). On the LMArena Chatbot Arena leaderboard, TurboS achieved a score of 1356, placing it in the top seven among models evaluated at the time, with top-five positions in Math, Multi-Turn, and Longer Query categories.

For Hy3 preview (April 2026): the Artificial Analysis Intelligence Index places Hy3 at a composite score of approximately 42, above the open-weights median of 29 among models of comparable active-parameter size. Tencent reported SWE-bench Verified performance of 74.4% and BrowseComp performance of 67.1%. On STEM-focused evaluations, Hy3 preview achieved strong results on the FrontierScience-Olympiad and the Tsinghua Qiuzhen College Math PhD qualifying exam (Spring 2026). Tencent characterized Hy3 as outperforming DeepSeek-V3 on GSM8K and MATH while using fewer total parameters.

Benchmark positions for all Hunyuan releases are point-in-time. The text LLM field rotates rapidly, and the gap between Hy3 and the Chinese open-weights frontier (DeepSeek V4, Qwen 3) is closer than headline parameter counts suggest.

Access and pricing

Open-weights releases for the Hunyuan text LLM line are distributed through the tencent organization on Hugging Face and the Tencent-Hunyuan GitHub organization. The Hunyuan-Large (Hunyuan-A52B-Pretrain, Hunyuan-A52B-Instruct, and Hunyuan-A52B-Instruct-FP8), Hunyuan-A13B (pretrain, instruct, FP8, GPTQ-Int4), and Hy3 preview weights are freely downloadable and compatible with the Hugging Face format for fine-tuning via hf-deepspeed and standard inference tooling.

The Tencent Cloud API at cloud.tencent.com provides access to Hunyuan 2.0 (Think and Instruct), Hunyuan-turbo, Hunyuan-pro, Hunyuan-standard, and Hunyuan-lite variants. Pricing increased in March 2026: HY2.0 Instruct input moved to approximately RMB 0.0045 per thousand tokens and output to approximately RMB 0.011 per thousand tokens. Hy3 preview via Tencent Cloud TokenHub is priced at RMB 1.2 per million input tokens and RMB 4 per million output tokens. First-time API users receive 1 million free tokens on Hunyuan-turbo.

Consumer access is through Tencent Yuanbao, which reached 41.64 million monthly active users by mid-2025 and topped the Apple App Store in China. In April 2025, WeChat launched a Yuanbao AI friend feature that allows users to add the assistant as a WeChat contact, extending Hunyuan's reach into WeChat's 1.4 billion monthly active user base.

Comparison

Direct competitors to the Hunyuan text LLM line as of May 2026:

• DeepSeek V4 (DeepSeek). The primary benchmark against which Hy3 preview was positioned at release. DeepSeek V4 Pro uses 1.6 trillion total parameters with 49 billion active, and holds an Artificial Analysis Intelligence Index composite of 51.51, roughly 9 points above Hy3's composite of 42. V4 Pro leads on the standardized benchmark set (SWE-bench rank 3, HumanEval+ rank 3, AIME rank 5). Tencent's counter is efficiency: Hy3 claims comparable MATH and coding performance at roughly one-fifth the total parameter count. Both are open-weights under permissive licenses.
• Qwen 3 (Alibaba). The closest Chinese open-weights peer on the text LLM dimension. Qwen 3 competes with Hy3 and DeepSeek V4 at the frontier of the Chinese open-weights ecosystem, with particular strength on Asian-language benchmarks. Hunyuan-TurboS's MMLU-Redux of 87.11 trails Qwen 3-235B-A22B's 87.40 on that benchmark, reflecting roughly parity rather than a clear hierarchy.
• Kimi K2 (Moonshot AI). A parallel Chinese open-weights MoE release that competes with Hy3 on agent and coding tasks. Hunyuan's differentiation against Kimi K2 rests on Tencent's consumer-distribution scale through Yuanbao and WeChat, which Moonshot AI as a standalone startup cannot match.
• Yi-Large (01.AI). Yi-Large represents an earlier generation of the Chinese open-weights text LLM frontier. As of May 2026, Hy3 and Hunyuan 2.0 both benchmark above the Yi-Large line on standard evaluations. The competitive framing has largely shifted from Hy3 vs. Yi to Hy3 vs. DeepSeek V4 and Qwen 3.

Hunyuan's distinct position in this comparison set: a text LLM line backed by Tencent's consumer-distribution infrastructure (WeChat, Yuanbao) and enterprise-cloud distribution (Tencent Cloud), combined with open-weights releases that keep the models competitive in developer and research contexts. No other model in this comparison set has equivalent consumer-product distribution.

Outlook

Open questions for the Hunyuan text LLM line over the next 6 to 18 months:

• Hy3 full release. Hy3 preview was released as a preview in April 2026; a production-grade full release with extended fine-tuning support and wider inference infrastructure would be the expected next step.
• Closing the benchmark gap with DeepSeek V4 Pro. Hy3's Artificial Analysis Intelligence Index composite of 42 trails V4 Pro's 51.51. Whether Tencent's rebuilt infrastructure delivers a successor that closes this gap, or whether the efficiency pitch (fewer active parameters for comparable task performance) becomes the durable competitive framing.
• Open-weights expansion. Tencent has open-sourced Hunyuan-Large, Hunyuan-A13B, and Hy3, but not Hunyuan 2.0. Whether the HY2.0 Think/Instruct weights follow, or whether the commercial API models remain proprietary, will affect Hunyuan's standing in the developer community.
• WeChat AI agent. Tencent's WeChat division began phased testing of an AI agent in mid-2026 that would integrate with WeChat's mini-program ecosystem and its 1.4 billion monthly active users. How the Hunyuan text LLM line powers that agent is the highest-stakes near-term consumer-AI decision.
• Tencent Cloud AI revenue growth. API price increases in March 2026 reflect rising demand and infrastructure costs. Whether Tencent Cloud can sustain developer and enterprise adoption alongside the price increases will be a metric to track through 2026.
• US export-control exposure. Tencent's AI infrastructure has not yet faced the same direct export-control attention as DeepSeek, but the regulatory environment for Chinese AI compute is uncertain through 2026 and 2027.

Sources

• Hugging Face: tencent/Tencent-Hunyuan-Large. Open-weights release page for Hunyuan-Large (Hunyuan-MoE-A52B).
• GitHub: Tencent-Hunyuan/Tencent-Hunyuan-Large. Model code, inference documentation, and November 2024 technical release notes.
• arXiv: Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent (2411.02265). Technical paper for the Hunyuan-Large release.
• Tencent: Hunyuan-T1 Release Notes. March 2025 release documentation and benchmark results for Hunyuan-T1.
• GitHub: Tencent-Hunyuan/Hunyuan-TurboS. TurboS release and technical notes for the Mamba-Transformer fast-thinking base.
• arXiv: Hunyuan-TurboS Technical Report (2505.15431). Benchmark and architecture detail for TurboS.
• GitHub: Tencent-Hunyuan/Hy3-preview. Hy3 preview release, architecture notes, and benchmark claims.
• Hugging Face: tencent/Hy3-preview. Open-weights distribution for Hy3 preview.
• Artificial Analysis: Hy3 Intelligence Index. Third-party composite benchmark score for Hy3 preview.
• TechNode: Tencent releases Hunyuan 2.0, its next-generation AI model. December 2025 Hunyuan 2.0 release coverage.
• TechNode: Tencent tests Yuanbao AI assistant within WeChat. March 2025 WeChat-Yuanbao integration reporting.