テンセントAI研究所오픈소스

Tencent HunYuan 3 Preview

Name: Tencent HunYuan 3 Preview
Author: テンセントAI研究所

이 모델 비교

텐센트 개발의 최신 기초 모델. 중국어 처리 및 멀티모달 대응에 강점이 있습니다.

공유:X Hatena LINE

파라미터

2950

컨텍스트

256K

라이선스

https://github.com/Tencent-Hunyuan/Hunyuan-A13B/blob/main/LICENSE

출시일

2026-04-23

벤치마크 성능

AA Intelligence Index

—

LMArena Elo

—

HLE

—

ARC-AGI-2

—

SWE-bench Verified

—

GPQA Diamond

—

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

일본어 처리 능력

🌐Multilingual

General multilingual model. Basic Japanese processing is possible, but inferior to specialized models.

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

약점

활용 사례

심층 분석

Total / Active Parameters

295B / 21B

MoE, 192 experts, top-8 activated — most efficient activation ratio in class

SWE-Bench Verified

74.4%

Coding agent benchmark

GPQA Diamond

87.2%

Graduate-level reasoning

MATH (4-shot, pretrained)

76.28

Best among Kimi-K2 (71.20), DeepSeek-V3 (59.37), GLM-4.5 (61.00)

Context Window

256K tokens

Native support

Input Price (Tencent Cloud)

~$0.18/1M tokens

~1/10th of GPT-4-class pricing; cached input ~$0.06/1M

Terminal-Bench 2.0

54.4

Search/reasoning agent benchmark

강점

・Exceptional parameter efficiency — 21B active parameters deliver performance rivaling models with 32–37B active, due to differentiated expert sizing and precision routing
・Aggressive pricing at ~$0.16–0.18/1M input tokens makes it one of the cheapest frontier-class APIs available, undercutting DeepSeek and Western models by an order of magnitude
・Deep integration with Tencent's massive consumer ecosystem (Yuanbao, WeChat, QQ — hundreds of millions of MAUs) provides unmatched distribution and real-world product feedback loops

약점

・Preview-stage release with known issues: weak error recovery during tool calls and sensitivity to inference hyperparameters; real-world stability not yet proven at scale
・PRC content filters apply across politically sensitive topics (Tiananmen, Taiwan sovereignty, Xi Jinping), matching DeepSeek/Qwen regulatory constraints
・English-language polish and nuanced writing lag behind Claude and GPT; best suited as a coding/reasoning/Chinese-language pick rather than an English prose-quality model

경쟁사 비교

Model	SWE	GPQA
Hy3 Preview (Tencent)	74.4%	87.2%
Kimi-K2 BASE (Moonshot)	N/A	N/A
DeepSeek-V3 BASE (DeepSeek)	N/A	N/A
GLM-4.5 BASE (Zhipu)	N/A	N/A

개요

Tencent Hy3 Preview (Hunyuan 3) is a 295-billion-parameter Mixture-of-Experts model with only 21 billion activated parameters, released and open-sourced on April 23, 2026. It represents the first model built on Tencent's entirely rebuilt pre-training and reinforcement learning infrastructure, led by Chief AI Scientist Shunyu Yao (formerly of OpenAI, known for the ReAct framework). The model integrates both fast and slow thinking in a single architecture, supporting configurable reasoning effort from direct response to deep chain-of-thought. Hy3 Preview's defining trait is its parameter efficiency: it achieves math and coding scores that match or exceed competitors with 50–75% more activated parameters (32–37B) and dramatically larger total parameter counts (671B–1,043B). This is enabled by differentiated expert sizing — routing tokens of varying difficulty to experts with different capacities — and a P-Penalty Loss that encourages activating more small experts. The 256K context window, Multi-Token Prediction (MTP) layer for speculative decoding, and three-tier reasoning effort control make it a flexible foundation for both consumer and enterprise agentic workflows. The model is already deeply integrated into Tencent's core products — Yuanbao (41.6M+ MAU), QQ, WeChat, CodeBuddy, WorkBuddy, and Tencent Docs — where it powers workflows of up to 495 steps. Pricing at ~$0.18 per million input tokens (approximately one-tenth of GPT-4-class models) positions Hy3 as a cost leader alongside DeepSeek. As a preview release, it carries known rough edges, but its open-weights availability under the Tencent Hy Community License, combined with Tencent's unparalleled consumer distribution, makes it a model to watch closely as the full release approaches.

벤치마크 및 성능

### Pre-trained Model Benchmarks (5-shot / 4-shot unless noted) | Benchmark | Hy3 Preview-Base (21B/295B) | Kimi-K2 BASE (32B/1043B) | DeepSeek-V3 BASE (37B/671B) | GLM-4.5 BASE (32B/355B) | |---|---|---|---|---| | **MMLU** | 87.42 | **88.24** | 87.68 | 87.73 | | **MMLU-Pro** | **65.76** | 65.98 | 63.98 | 63.67 | | **SuperGPQA** | **51.60** | 51.10 | 46.17 | 49.64 | | **MATH** (4-shot) | **76.28** | 71.20 | 59.37 | 61.00 | | **GSM8K** (4-shot) | **95.37** | 93.46 | 88.15 | 90.06 | | **CMath** (4-shot) | **91.17** | 90.83 | 85.50 | 89.33 | | **LiveCodeBench-v6** (1-shot) | **34.86** | 30.86 | 29.31 | 27.43 | | **CRUXEval-I** (3-shot) | **71.19** | 68.01 | 67.79 | 68.51 | | **MMMLU** (5-shot) | **80.15** | 77.63 | 79.54 | 79.26 | | **INCLUDE** (5-shot) | **78.64** | 75.66 | 77.86 | 76.27 | | **ARC-Challenge** (0-shot) | 95.99 | **96.66** | 94.65 | 96.32 | | **DROP** (5-shot) | 85.50 | 86.40 | **86.50** | 82.90 | ### Instruct Model Benchmarks | Benchmark | Hy3 Preview | Notes | |---|---|---| | **SWE-Bench Verified** | **74.4%** | Coding agent benchmark | | **GPQA Diamond** | **87.2%** | Graduate-level science QA | | **Terminal-Bench 2.0** | **54.4** | Long-horizon terminal agent | | **HLE (Harbor Framework)** | **30.0** | Expert-level evaluation | ### Key Takeaways - Hy3 Preview-Base leads on **7 of 12** pre-trained benchmarks despite having the **fewest active parameters** (21B vs 32–37B) - The MATH score of 76.28 is **5+ points above** the next-best competitor (Kimi-K2 at 71.20) - LiveCodeBench-v6 at 34.86 is **4 points above** Kimi-K2 and **12+ points above** GLM-4.5 - Multilingual (MMMLU, INCLUDE) scores are uniformly best-in-class, suggesting strong cross-lingual transfer - English knowledge benchmarks (MMLU, ARC) show Hy3 trailing slightly, indicating room for improvement in general English knowledge retrieval

상세 비교

### Hy3 Preview vs DeepSeek-V3 / V4 | Dimension | Hy3 Preview | DeepSeek-V3 | DeepSeek V4 Pro (reference) | |---|---|---|---| | **Architecture** | MoE, 295B total / 21B active | MoE, 671B total / 37B active | MoE, ~1.6T total / 49B active | | **Context** | 256K | 128K | 128K+ | | **MATH (pretrained)** | 76.28 | 59.37 | N/A | | **Input Price** | ~$0.18/1M | ~$0.27/1M | ~$0.27/1M | | **SWE-Bench Verified** | 74.4% | N/A (V3) | Higher (V4 Pro) | | **License** | Tencent Hy Community | MIT | MIT | | **Key Strength** | Parameter efficiency, math | Ecosystem maturity, MoE scaling | Higher ceiling, broader benchmarks | | **Key Weakness** | Preview stage, English prose | Smaller context (128K) | Higher cost | DeepSeek-V3 serves as the primary benchmark competitor. Hy3 outperforms it on all pre-trained benchmarks while using 46% fewer active parameters and 56% fewer total parameters. However, DeepSeek V4 Pro (not directly tested against Hy3 in the same table) achieves a higher Artificial Analysis Intelligence Index composite (~51.5 vs Hy3's ~42), suggesting the gap closes at the true frontier. DeepSeek's MIT license is more permissive than Tencent's community license. ### Hy3 Preview vs Kimi-K2 (Moonshot AI) | Dimension | Hy3 Preview | Kimi-K2 | |---|---|---| | **Architecture** | MoE, 295B / 21B active | MoE, 1043B / 32B active | | **MATH** | 76.28 | 71.20 | | **GSM8K** | 95.37 | 93.46 | | **MMLU** | 87.42 | 88.24 | | **LiveCodeBench** | 34.86 | 30.86 | Kimi-K2 uses 3.5× more total parameters and 52% more active parameters but trails Hy3 on math, code, and multilingual benchmarks. Kimi-K2 edges ahead on MMLU (+0.82 points) and SimpleQA (+7.9 points), suggesting marginally stronger English knowledge retrieval. Tencent's distribution moat (WeChat, QQ) is a structural advantage Moonshot cannot match. ### Hy3 Preview vs GLM-4.5 (Zhipu AI) GLM-4.5 uses 32B active out of 355B total. Hy3 outperforms it on every benchmark in the comparison table except ARC-Challenge (96.32 vs 95.99). The efficiency gap is notable: Hy3 uses 34% fewer active parameters to achieve superior results. GLM-4.5's smaller total parameter count (355B vs 295B) means its architecture is more compact, but Hy3's MoE routing appears more effective.

커뮤니티 평가

The developer and research community has responded to Hy3 Preview with cautious optimism. The HuggingFace repository (tencent/Hy3-preview) accumulated 78,998 downloads in its first month and 431 GitHub stars within days of release, indicating strong initial interest. Key community reactions: - **Efficiency praise**: Multiple analysts highlight the 21B active parameter efficiency as genuinely differentiated. The HuggingFace community article by Leco Li emphasizes that 'what others achieve with 37B–40B activated parameters, Hunyuan accomplishes with just 21B,' attributing this to differentiated expert sizing and routing precision. - **Pricing enthusiasm**: The ~$0.18/1M input pricing has been widely noted as aggressive. AIToolTier rates the model's Value score at 9.5/10, calling it 'among the cheapest frontier-class APIs on the market.' The SiliconFlow provider offers even lower rates ($0.066/1M input), making it accessible for experimentation. - **Preview-stage caution**: Multiple reviewers note the preview status. AIToolTier states: 'treat Tencent's own benchmark claims with the usual self-reporting discount until independent runs corroborate.' HTX Insights' hands-on testing found the model 'steady but not stunning,' with trap identification and complex open-ended task completion showing room for improvement. - **Distribution recognition**: The integration into WeChat (1.4B MAU) and Yuanbao (41.6M+ MAU) is widely seen as a structural moat. Nextomoro notes: 'No other model in this comparison set has equivalent consumer-product distribution.' - **Agent workflow traction**: The model has been integrated with popular open-source agent frameworks (OpenClaw, OpenCode, KiloCode) and Tencent reports it can reliably power workflows of up to 495 steps in production environments. The 54% reduction in Time-To-First-Token and 47% reduction in end-to-end latency within Tencent's CodeBuddy and WorkBuddy products are cited as practical validation. - **License concerns**: The Tencent Hy Community License is noted as less permissive than Apache 2.0 or MIT, with potential commercial-scale restrictions that may concern some developers. LLMPick flags: 'community license has commercial-scale restrictions.' Overall sentiment: Hy3 is viewed as a strong efficiency play and a credible sign that Tencent is finally competitive in the open-weights race, but the community is waiting for the full GA release and more independent benchmark verification before treating it as a proven production option.

활용 사례

### 1. Cost-Sensitive Chinese Market Applications **When to choose Hy3 over alternatives**: For applications targeting Chinese-speaking users where cost-per-token is critical, Hy3 is the clear leader. At ~¥1.2 RMB/1M input tokens, it undercuts DeepSeek-V3 and approaches the pricing of smaller, less capable models. Example: A Chinese e-commerce chatbot handling millions of daily customer queries across WeChat mini-programs can leverage Hy3's native Chinese language performance (C-Eval 89.80, CMMLU 89.61) while keeping inference costs manageable. ### 2. Coding Agent and Software Engineering Workflows **When to choose Hy3 over alternatives**: For multi-step coding tasks in development environments, Hy3's SWE-Bench Verified score of 74.4% and Terminal-Bench 2.0 score of 54.4% make it competitive with far more expensive models. The three-tier reasoning effort control (no_think / low / high) is particularly valuable: use 'high' for complex debugging and architecture tasks, 'no_think' for quick code completions. Tencent's CodeBuddy product demonstrates this pattern, achieving 99.99%+ success rate with 47% reduced end-to-end latency. ### 3. Long-Context Document Processing and Information Extraction **When to choose Hy3 over alternatives**: For tasks requiring processing of lengthy documents (meeting transcripts, legal contracts, research papers) up to 256K tokens, Hy3's Context Learning capabilities — validated on Tencent's proprietary CL-bench benchmarks — make it suitable for enterprise document workflows. Example: Processing 100+ page technical manuals to extract structured information and generate interactive content (as demonstrated in the WorkBuddy textbook-to-quiz-game scenario). Choose Hy3 over DeepSeek-V3 when you need 256K context (vs 128K) at lower cost. ### 4. Complex Multi-Step Agent Orchestration **When to choose Hy3 over alternatives**: For agent workflows that require tool calling, web search, and multi-step reasoning chains, Hy3 has demonstrated reliable execution of workflows up to 495 steps in production. Its integration with OpenClaw, OpenCode, and KiloCode frameworks, combined with MCP (Model Context Protocol) toolchain support, makes it suitable for building autonomous research agents, data analysis pipelines, and document processing chains. Choose Hy3 over alternatives when you need strong agent capabilities at frontier-class quality but cannot justify the cost of GPT-4-class APIs. **When NOT to choose Hy3**: For English-first creative writing, nuanced prose, or politically sensitive content involving PRC-regulated topics, Claude or GPT models remain stronger choices. For use cases requiring maximum benchmark ceiling (not efficiency), DeepSeek V4 Pro or Claude Fable 5 may be more appropriate.