Hy3

The Hugging Face model card for tencent/Hy3-preview documents the official open-weights release of Hy3 preview, a 295B-parameter Mixture-of-Experts model with 21B active parameters, 3.8B MTP layer parameters, 192 experts with top-8 routing, 80 layers, 64 attention heads (GQA, 8 KV heads, head dim 128), hidden size 4096 According to the same Hugging Face page, Hy3 preview highlights gains in STEM and reasoning benchmarks including FrontierScience-Olympiad, IMOAnswerBench, the Tsinghua Qiuzhen College Math PhD qualifying exam (Spring '26), and the China High School Biology Olympiad (CHSBO 2025). Context learning and instruction followi

Tencent announced broader international access to Hy3 on August 5, 2026, following its official July 6 release, under the unified Tencent Hy branding (formerly Tencent Hunyuan). Users and enterprises can access the model globally through WorkBuddy free of charge until August 31, 2026 (Pacific Time), as well as through Tencent reported that Hy3 recorded more than 68 times as many API calls as the previous-generation model and ranked first globally on OpenRouter's usage leaderboard within one week of launch. The model is distributed across third-party developer platforms including Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry St

Tencent announced broader international access to Hy3 on August 5, 2026, after the model’s official July 6 release. Developers can connect through an API, while users and enterprises can access it through WorkBuddy, Tencent Design Miora, and Tencent Cloud TokenHub; further integrations across Tencent Cloud’s AI-native The announcement describes Hy3 as a hybrid fast-and-slow-thinking Mixture-of-Experts model with 295 billion total parameters, 21 billion active parameters, and context windows up to 256K. Tencent also reports more than 68 times the API calls of its previous-generation model and a number-one OpenRouter global LLM usage

MarkTechPost's coverage (published July 7, 2026) confirms Tencent's Hy team released Hy3 as a 295B-parameter Mixture-of-Experts model activating only 21B parameters per token, with weights released under Apache License 2.0. The architecture uses a sparse MoE with 192 experts and top-8 routing, plus a Multi-Token Predic The same MarkTechPost article targets reasoning, agentic workflows, and long-context tasks, and lists the full specification table: 80 layers excluding MTP, 1 MTP layer, 64 attention heads (GQA with 8 KV heads and head dim 128), hidden size 4096, intermediate size 13312, 256K context length, and vocabulary size 120832.

LLM Stats tracks Hy3 with an overall rank of 41 on its composite LLM Stats Score, placing it in the B average capability tier. Category-specific standings include Search rank 12 of 90, Long Context rank 14 of 119, Math rank 33 of 327, Coding rank 35 of 267, Reasoning rank 42 of 363, and Tool Calling rank 50 of 194, sit Individual benchmark scores include USAMO 2026 at 30.24/42 (rank 3), DeepSearchQA at 0.91/1 (rank 5), GPQA Diamond at 0.90/1 (rank 25), and IMO-AnswerBench at 0.90/1 (rank 3), with 27 additional benchmark results tracked on the page. The site notes these scores are self-reported by the model provider and may not be ind

Kilo announced General Availability of Tencent's Hy3 on July 6, 2026, making it freely available across its VS Code extension, cloud agents, and CLI with high limits thanks to compute from Novita. The GA release scaled up reinforcement learning training and refined post-training data quality, with Kilo reporting that H Documented GA improvements include production-grade tool calling with better success rates and error recovery, a drop in internal hallucination rate from 12.5% to 5.4% after data cleaning and training constraints, and upgraded multi-turn long-context intent tracking for complex, multi-step refactoring sessions. Kilo po

Artificial Analysis' independent evaluation page (Tencent, open weights, released July 2026) places Hy3 at an Intelligence Index of 25, ranking it well above the median of 18 among comparable models, with 86.2 output tokens per second and 170M tokens generated during the Intelligence Index evaluation. The page records The same Artificial Analysis page lists Hy3 at 299B total parameters with 21B active parameters per token, under Apache 2.0, available via Hugging Face, and flagged as the reasoning variant with a note that a non-reasoning variant may also exist. A minor parameter-count discrepancy versus Tencent's 295B figure is noted

Tencent's Hunyuan team open-sourced the Hy3 preview on April 23, 2026, after a reorganization, releasing it as a Mixture-of-Experts model with 295B total parameters, 21B activated parameters, and a 256K context window. The model was deployed across Tencent products including Tencent Cloud, Yuanbao, QQ, and Tencent Docs By April 29, 2026, the OpenRouter global large model API call volume ranking placed Hy3 preview first overall, first in tool calling, and second in programming scenarios. Tencent officially released the production Hy3 on July 6, 2026, with further improvements in complex reasoning, instruction following, and code gener