Hy3 - Tencent Hunyuan's open-source hybrid expert model

What is Hy3?

Hy3 is a 295-parameter hybrid expert (MoE) model open-sourced by Tencent Hunyuan. The model shows significant improvements in agent capabilities, inference, and long-context tasks, achieving benchmark performance comparable to flagship models with 2-5 times the parameter size in multiple benchmarks. Through enhanced training and data cleaning, the illusion error rate is reduced to 5.4%, and the commonsense error rate is reduced to 12.7%. Hy3 supports diverse productivity scenarios such as code development, office automation, and front-end design, making it a reliable and cost-effective choice.

Hy3's main functions

• Intelligent agent taskSignificant progress has been made in tasks such as reasoning, agents, and long contexts, with multiple benchmarks comparable to flagship models with 2-5 times the parameter size.
• Code developmentIt supports software development, front-end design, CI/CD, etc., and performs excellently on the SWE-bench series benchmarks.
• Office productionSupports Excel modeling and analysis, PPT creation, complex data summarization, and linked formula calculation.
• Tool callEnhanced tool call stability and cross-scaffolding generalization, supporting Codebuddy, Cline, KiloCode, etc.
• Anti-hallucinationBy using fine-grained data cleaning and training constraints, the rate of hallucinations and common sense errors can be significantly reduced.
• Multiple rounds of dialogueOptimize referential resolution, omission restoration, and multi-round constraint inheritance to maintain consistency of intent in complex long-term interactions.

Hy3's technical principles

• Post-training optimizationBased on the preview version, improve the quality and diversity of training data to expand the scale of RL computing power.
• Joint trainingJointly optimize business pain points such as referential resolution, omission restoration, and multi-round constraint inheritance in the SFT and RL stages.
• Data cleaningWe will perform fine-grained data cleaning and training constraints, following the principle of "answering only when there is evidence and explicitly stating the lack of evidence when there is none".
• Hardware and software synergyReduce inference costs and improve token efficiency through hardware and software co-optimization.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use Hy3

• Obtain model weightsVisit GitHub or HuggingFace to download the model weights and code.
• Activate API serviceYou can activate the Hy3 API service in the Tencent Cloud console and obtain the access key to make API calls.
• Configure the runtime environmentConfigure the runtime environment locally or in the cloud, load the model, and connect to scaffolding tools such as Codebuddy, Cline, and KiloCode.
• Call to complete task: Call the model via API or local deployment to complete tasks such as code development, office automation, and data analysis.

Hy3's core advantages

• Performance exceeding expectationsIt boasts significantly superior intelligence compared to other models of the same size, and its performance rivals that of flagship models with 2-5 times the number of parameters in multiple benchmarks.
• Reliable experienceThe model illusion rate decreased from 12.5% to 5.4%, the common sense error rate decreased from 25.4% to 12.7%, and the multi-turn question rate decreased from 17.4% to 7.9%.
• Tool stabilityThe standard deviation of the SWE Bench Verified score across different scaffolding structures was controlled within 4 percentage points, and the tool call error recovery capability was greatly improved.
• Low costAPI input costs 1 yuan per million tokens, output costs 4 yuan per million tokens, and cache hits cost only 0.25 yuan per input. Token consumption for office tasks is reduced by approximately 47-49% compared to competitors.

Hy3's project address

• Project official websitehttps://hy.tencent.com/research/hy3
• GitHub repositoryhttps://github.com/Tencent-Hunyuan/Hy3
• HuggingFace model libraryhttps://huggingface.co/tencent/Hy3

Hy3's comparison with similar competing products

Hy3 application scenarios

• Software developmentIt supports code generation, debugging, and CI/CD process automation, and performs excellently on benchmarks such as SWE-bench.
• Office AutomationSupports Excel modeling and analysis, PPT creation, complex data summarization, and linked formula calculation.
• Front-end designIt can generate front-end code such as directly executable HTML web pages and browser plugins.
• Intelligent Customer ServiceIt supports AI avatars and customer service scenarios, and can handle incomplete expressions and make reasonable judgments based on context.
• Game developmentSupports game logic writing and game assistant integration, such as the AI assistant in Path of Exile: Advent.