이 모델의 강점은 무엇인가요?

방대한 2.95조 파라미터 256K 긴 컨텍스트 지원 고급 추론 능력을 고려하여 설계됨

이 모델의 약점은 무엇인가요?

거대한 모델로 인한 높은 계산 부담 제한된 사용 라이선스 프리뷰 버전으로 인한 불완전한 기능

어떤 용도에 가장 적합한가요?

복잡한 논리적 추론 작업 대규모 문서 세트의 분석 및 요약 고급 지식 베이스 구축

모델 목록으로

テンセントAI研究所조건부 오픈

Tencent HunYuan 3 Preview

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

Tencent Hunyuan 3 Preview는 Tencent AI Lab이 개발한 대규모 추론 모델입니다. 약 2.95조(2.95T) 파라미터의 초대규모 스케일을 가지고 있으며, 넓은 256K 컨텍스트 윈도우를 지원합니다.

파라미터

2950.0B

컨텍스트

256K

라이선스

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

출시일

2026-04-23

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

・방대한 2.95조 파라미터
・256K 긴 컨텍스트 지원
・고급 추론 능력을 고려하여 설계됨

약점

・거대한 모델로 인한 높은 계산 부담
・제한된 사용 라이선스
・프리뷰 버전으로 인한 불완전한 기능

활용 사례

・복잡한 논리적 추론 작업
・대규모 문서 세트의 분석 및 요약
・고급 지식 베이스 구축

심층 분석

Total Parameters

295B

Mixture-of-Experts (MoE) Architecture

Active Parameters

21B

Optimized for cost-efficiency

Context Window

256K tokens

Supports long-context processing

MATH Benchmark (4-shot)

76.28

Pre-trained model, best among peers

API Input Price

¥1.2 / million tokens

~$0.17, aggressive pricing on Tencent Cloud

Deployment Scale

8x H20-3e GPUs

Recommended for self-hosting

강점

・Exceptional STEM reasoning and math performance, outperforming DeepSeek-V3 on key benchmarks with far fewer total parameters.
・Pioneering 'fast & slow thinking' fusion in a single model, with configurable reasoning effort for complex agent tasks.
・Deeply integrated into Tencent's product ecosystem (Yuanbao, WorkBuddy) and optimized for real-world agent workflows of up to 495 steps.

약점

・As a preview model, it exhibits weak error recovery during tool calls and is sensitive to inference hyperparameters.
・Internal tests show instability in long-chain agent execution, with tasks sometimes being only partially completed.
・No direct API access from major global cloud providers (AWS, GCP, Azure) yet, requiring self-hosting or use of Tencent Cloud/OpenRouter.

경쟁사 비교

Model	Arena	SWE	GPQA	Price
DeepSeek-V3	N/A	N/A	N/A	~$0.27/$1.10 (DeepSeek API)
Kimi-K2	N/A	N/A	N/A	Approx. similar tier
GLM-4.5	N/A	N/A	N/A	Tiered pricing on Zhipu API

개요

Tencent Hunyuan 3 (Hy3) Preview is Tencent's most powerful open-weights large language model to date, released in April 2026. It represents a strategic pivot for Tencent's AI efforts, rebuilt from the ground up in 90 days under the leadership of Chief AI Scientist Yao Shunyu. As a 295-billion-parameter Mixture-of-Experts model with only 21 billion active parameters, Hy3 prioritizes practical utility and cost-efficiency over sheer scale, aiming to deliver performance competitive with larger models at a fraction of the operational cost. The model's core innovation is its 'fusion reasoning' architecture, which integrates both 'fast' (direct response) and 'slow' (deep chain-of-thought) thinking modes into a single system, controllable via an API parameter. It shows significant strengths in complex STEM reasoning, coding, and—most notably—agentic tasks. Tencent has tightly co-designed the model with its own product ecosystem, including the Yuanbao chatbot and WorkBuddy AI assistant, and has demonstrated its ability to power stable, multi-hundred-step agent workflows in real enterprise scenarios. Positioned as a cost-efficient powerhouse in the competitive Chinese open-weights landscape, Hy3 directly competes with DeepSeek-V3 and Qwen 3. Its release is coupled with aggressive API pricing on Tencent Cloud and immediate integration into Tencent's vast product portfolio, leveraging the company's unparalleled distribution through WeChat and QQ to drive adoption. It is a preview release, signaling more refined and capable versions to come.

벤치마크 및 성능

Hy3 Preview demonstrates leading performance among comparable open-weights models, particularly in math and coding, despite having significantly fewer total parameters than competitors. The benchmarks below compare the *pre-trained base models*. | Benchmark (Metric) | Kimi-K2 Base (1043B) | DeepSeek-V3 Base (671B) | GLM-4.5 Base (355B) | **Hy3 Preview-Base (295B)** | | :--- | :---: | :---: | :---: | :---: | | **MMLU (5-shot)** | **88.24** | 87.68 | 87.73 | 87.42 | | **MATH (4-shot)** | 71.20 | 59.37 | 61.00 | **76.28** | | **GSM8K (4-shot)** | 93.46 | 88.15 | 90.06 | **95.37** | | **LiveCodeBench-v6 (1-shot)** | 30.86 | 29.31 | 27.43 | **34.86** | | **MMMLU (5-shot)** | 77.63 | 79.54 | 79.26 | **80.15** | | **INCLUDE (5-shot)** | 75.66 | 77.86 | 76.27 | **78.64** | *For the instruction-tuned model, Tencent reports strong results on agent-focused benchmarks:* * **SWE-bench Verified (Coding Agent):** 74.4% * **Terminal-Bench 2.0 (Terminal Agent):** 54.4 * **BrowseComp (Search Agent):** 67.1% *Note: Direct scores for Arena Elo, GPQA, and SWE-Bench for the direct competitors (DeepSeek-V3, etc.) are not provided in the source materials for a direct side-by-side, limiting a complete quantitative comparison table.*

상세 비교

**1. vs. DeepSeek-V3:** * **Positioning:** DeepSeek-V3 (405B active parameters in its dense variant) is the current open-weights king on raw reasoning benchmarks. Hy3 Preview is Tencent's efficiency-first challenger. * **Performance:** Hy3's pre-trained base beats DeepSeek-V3's base on critical math and multilingual benchmarks (MATH: 76.28 vs. 59.37, MMMLU: 80.15 vs. 79.54) with less than half the total parameters. * **Cost:** Hy3's API pricing on Tencent Cloud (~$0.17 input/M tokens) is highly competitive, aiming to undercut established players. * **Strengths:** DeepSeek likely retains an edge on absolute knowledge and some pure reasoning tasks. Hy3's advantage is its fusion reasoning and deep product integration for agents. **2. vs. Kimi-K2:** * **Positioning:** Both are Chinese MoE models focused on practical applications. Kimi-K2 has a larger total parameter count (1043B). * **Performance:** The models are closely matched on many English benchmarks (MMLU). Hy3 shows a clear lead in math (MATH, GSM8K) and multilingual (MMMLU, INCLUDE) tasks. * **Ecosystem:** Hy3 benefits from immediate, massive distribution through Tencent's consumer products (Yuanbao, WeChat), an advantage Kimi lacks. **3. vs. Qwen 3:** * **Positioning:** The primary domestic rivalry. Both are seen as the top tier of Chinese open-weights MoE models. * **Performance:** Performance is comparable on common benchmarks, with each model showing strengths in different areas. Hy3's claim is stronger efficiency (fewer activated parameters for similar results). * **Strategy:** Qwen focuses on a broad model family and developer ecosystem. Tencent leverages Hy3 to power and enhance its own vast product suite, creating a unique 'co-design' advantage.

커뮤니티 평가

Developer and industry reaction to Hy3 Preview has been a mix of strategic interest and cautious evaluation: * **Technical Focus:** Articles from 36Kr and Hugging Face blogs highlight the architectural novelty of the 'fusion reasoning' MoE and the significance of achieving strong math/coding scores with only 21B active parameters. The 90-day rebuild timeline is noted as impressive. * **Practical Skepticism:** Early hands-on tests (like those from 36Kr's GeekPark) show the model has clear agentic capabilities but is not yet fully stable for complex, real-world automation, often getting stuck or delivering partial results—consistent with Tencent's 'preview' label. * **Strategic Importance:** The release is widely viewed as Tencent's major move to reclaim its position in the AI race, driven by the high-profile hire of Yao Shunyu. The focus on product integration over pure leaderboard chasing is a noted strategic shift. * **Adoption:** Early adoption is being driven by its integration into Tencent's ecosystem and availability on OpenRouter. Its standing in the broader developer community will depend on the stability and features of the upcoming official release.

활용 사례

1. **AI-Powered Coding & DevOps Assistant (CodeBuddy):** Hy3 is integrated into Tencent's CodeBuddy, where it assists with code generation, debugging, and automation. Its configurable 'high' reasoning mode allows it to tackle complex multi-file tasks and infrastructure-as-code problems. *Choose Hy3 over alternatives when you need a model deeply integrated into an IDE-like environment with support for long-context, multi-step coding workflows and are operating within or targeting the Tencent ecosystem.* 2. **Complex Enterprise Agent Workflows (WorkBuddy):** Hy3 powers WorkBuddy to handle multi-hundred-step office automation: parsing long reports, executing data analysis via Python, coordinating tool calls (email, calendar, docs), and generating structured outputs. Its long context window and instruction following are critical here. *Choose Hy3 for internal enterprise tools where agents must follow intricate company-specific rules and operate on lengthy internal documents, especially if leveraging Tencent Cloud's infrastructure.* 3. **Contextual Learning and Knowledge Retrieval:** Using its dedicated 'context learning' capabilities, Hy3 can be used to build systems that learn from and act upon very large provided contexts (e.g., entire manuals, legal documents, or codebases). It can generate quizzes, extract structured data, or answer complex questions grounded in the provided text. *Choose Hy3 for applications requiring high-fidelity information extraction and instruction following over large, unstructured reference material provided at inference time.* 4. **Cost-Efficient STEM Reasoning & Tutoring:** With its leading performance on math benchmarks (MATH, GSM8K) and configurable reasoning depth, Hy3 can serve as the core engine for a low-cost, high-capability AI tutor for mathematics and sciences. *Choose Hy3 over larger, more expensive models when building applications where mathematical reasoning and step-by-step explanation quality are paramount, and cost-per-query is a major constraint.*