Tencent Releases and Open-Sources Tencent Hy4 Preview

Ranked among the apical tier of open-source models, Hy4 preview is built for real-world productivity tasks, delivering outstanding capacity crossed coding, agency work, and technological research

Tencent has released and open-sourced Tencent Hy4 preview, a next-generation ample connection exemplary pinch 770B full parameters and 49B progressive parameters, and a discourse model exceeding 1M tokens. It demonstrates outstanding capabilities connected real-world productivity tasks spanning coding, agency work, and technological research.

Hy4 preview is now disposable arsenic an open-source exemplary and tin besides beryllium accessed globally done WorkBuddy and CodeBuddy, arsenic good arsenic Yuanbao, ima and different Tencent products. Users tin effort the exemplary straight done these applications, aliases link to it via API done Tencent Cloud TokenHub and OpenRouter.

Upon launch, Hy4 preview will beryllium disposable for free connected WorkBuddy and CodeBuddy for 2 weeks.access to Hy3 connected some platforms has besides been extended until September 30.

Hy4 preview was expanded importantly successful exemplary size, discourse length, and information volume, and  the advances successful some pre-training and post-training person led to a awesome leap successful wide intelligence, placing the exemplary among the apical tier of open-source models.

Hunyuan continuously useful successful heavy co-design pinch products specified arsenic CodeBuddy and WorkBuddy, optimizing the real-world personification acquisition crossed productivity scenarios. In a unsighted information conducted internally by Tencent involving 163 experts and 203 engineering tasks, Hy4 preview scored an mean of 2.99 retired of 4.00, somewhat up of GLM-5.3 (2.92/4.00) and Kimi K3 (2.94/4.00).

Designed for productivity, Hy4 preview was developed utilizing high-quality training information co-created pinch Tencent experts crossed package engineering, gaming, finance, security, and different domains, arsenic good arsenic done heavy co-design pinch products specified arsenic WorkBuddy. This has helped thrust important improvements crossed a wide scope of real-world productivity tasks.

In package engineering, Hy4 preview delivers stronger understanding, planning, debugging, and validation capabilities for long-context improvement tasks, while besides enhancing the ocular value and relationship acquisition of front-end development.

In agency productivity and analytical scenarios, the exemplary demonstrates a importantly stronger knowing of analyzable moving environments and enhanced financial study capabilities. It has besides been optimized for information study and cross-document collaboration, supporting the afloat workflow from accusation processing done to the creation of documents, spreadsheets, and presentations.

In crippled development, Hy4 preview tin make a playable prototype from a azygous natural-language request, and activity efficaciously pinch crippled engines. Developers tin past proceed refining analyzable crippled projects done multi-turn interactions.

In technological research, Hy4 preview demonstrates stronger capabilities successful understanding, reasoning done and solving analyzable investigation problems, pinch notable improvements crossed areas including AI investigation and development, molecular dynamics simulation, condensed-matter physics and basal mathematics.

Notably, Hy4 preview besides contributed to its ain improvement process, participating for the first clip successful the automated optimization of training methods, information strategies, information frameworks, and low-level operators. The exemplary projected approaches, ran experiments, and iterated based connected the results, pinch the resulting code, logs, and feedback feeding into consequent rounds of exploration. This established an early-stage recursive self-improvement loop.

Hy4 preview has besides autonomously analyzed bottlenecks successful its conclusion system  and carried retired aggregate rounds of optimization connected areas specified arsenic usability fusion and connection optimization. These improvements accrued end-to-end throughput by 31.8% compared pinch the baseline, pinch accordant gains crossed different discourse lengths and concurrency levels. This demonstrates the model’s expertise to autonomously optimize its ain conclusion infrastructure.

Hy4 preview continues to connection costs efficiency, helping make precocious AI much wide accessible. API pricing is group astatine USD 0.834 per cardinal input tokens, USD 2.501 per cardinal output tokens and USD 0.042 per cardinal tokens for cache hits.

Through a preview-first approach, followed by charismatic releases, Hunyuan continuously incorporates real-world feedback into its investigation and improvement process, enabling its models to amended by solving real-world problems. The adjacent batch of models successful the Hy4 bid is expected to rotation retired soon.