Hy3

The Hugging Face model card confirms Hy3 preview is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team, and the first model trained on their rebuilt infrastructure. Detailed architecture is provided: 80 layers excluding MTP plus 1 M Hy3 preview is described as the strongest model Tencent has shipped so far, with notable gains on STEM and reasoning benchmarks (FrontierScience-Olympiad, IMOAnswerBench, the Tsinghua Qiuzhen College Math PhD qualifying exam Spring 2026, and CHSBO 2025), on context learning and instruction following using Tencent-built

An independent price-comparison page listed Hy3 preview across six providers as of early September 2026, with the cheapest listed input price at $0.066 per million tokens on SiliconFlow and the cheapest listed output price at $0.26 per million tokens on the same provider, though SiliconFlow caps context at 262K. ZenMux The aggregator recorded its 66th provider-coverage snapshot on September 1, 2026, with 12 daily snapshots between August 21 and September 1 showing no movement in best input or output prices. A changelog showed context window reductions on Kilo Gateway and OpenRouter on August 25, dropping output limits from 262K to 23

Tencent announced broader international access to its Hy3 large model on August 5, 2026, following the model's official release on July 6, 2026. The rollout provides global developers, users, and enterprises multiple integration paths, including the WorkBuddy application, Tencent Design Miora, and Tencent Cloud TokenHu Hy3 is built on a hybrid fast-and-slow-thinking Mixture-of-Experts architecture with 295 billion total parameters, 21 billion active parameters, and a context length of up to 256K. According to Tencent, the model performs comparably to flagship models with two to five times as many parameters, recorded more than 68 tim

VentureBeat's July 6, 2026 coverage documents the full Hy3 release from Tencent's Hunyuan team, a 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters, top-8 routing across 192 experts, a 3.8-billion-parameter multi-token prediction layer, and a 256K context window. A central technical and l The article reports that the full Hy3 was finalized roughly ten weeks after the preview through feedback from more than 50 internal product teams, with Tencent's post-training pipeline scaled and issues in task execution and interaction fixed while the MoE architecture stayed unchanged. It positions Hy3 as a reliabilit

Third-party benchmark aggregator llm-stats ranks Hy3 at position 40 on its composite LLM Stats Score with a value of 43.2 and a blended price of $0.16 per million tokens. The model sits in the B average capability tier overall, placing in the top half for Search (11 of 89), Long Context (14 of 119), and within range fo Detailed benchmark scores for Hy3 include rank 3 on USAMO 2026 with 30.24 out of 42, rank 4 on DeepSearchQA at 0.91 out of 1, rank 25 on GPQA Diamond at 0.90 out of 1, and rank 3 on IMO-AnswerBench at 0.90 out of 1, with sources pointing to huggingface.co. The aggregator notes that scores are self-reported by the model

On 2026-04-23, Tencent Hunyuan released and open-sourced the Hy3 preview language model, described as a hybrid fast-and-slow-thinking Mixture-of-Experts (MoE) model with 295B total parameters, 21B active parameters, and up to 256K context. The article, dated 2026-07-27, frames Hy3 preview as the first step of Hunyuan's The piece confirms the preview is open source and lists key specifications alongside deployment options on Tencent Cloud. The source supplies first-party confirmation of the architecture, parameter counts, context length, and the rebuilt pretraining and reinforcement learning infrastructure work that began in February,