Hy3

The Hugging Face model card for tencent/Hy3-preview confirms the model is a Mixture-of-Experts architecture with 295B total parameters, 21B activated parameters, and 3.8B MTP layer parameters, running on Tencent's rebuilt training infrastructure. It supports a 256K context window, uses GQA with 8 KV heads at a head dim Deployment quickstarts for vLLM and SGLang are included directly on the model card, making the model immediately runnable for self-hosters and giving OrcaRouter a concrete open-weights foundation for its routing endpoint. The preview status is explicitly noted, which is relevant context for any production-grade integra

Tencent announced broader international access to the officially released Hy3 model on August 5, 2026, following its July 6 launch, making it available through WorkBuddy (free to global users until August 31, 2026 Pacific Time), Tencent Design Miora, and Tencent Cloud TokenHub, with developer API access and further clo The release reiterates Hy3's core technical profile as a hybrid fast-and-slow-thinking Mixture-of-Experts model with 295 billion total parameters, 21 billion active parameters, and a 256K context window, and claims it performs comparably to flagship models with two to five times as many parameters. It also reports that

Tencent officially released its Hy3 open-source AI model on July 6, 2026, featuring a Mixture-of-Experts architecture with 295 billion total parameters, 21 billion active parameters, an additional 3.8 billion MTP layer parameters, and a 256,000-token context window. In a blind evaluation by 270 experts, Hy3 scored 2.67 Hy3 is available under an Apache 2.0 license on Hugging Face, ModelScope, and GitHub, with an FP8-quantized variant also offered. Tencent has already integrated Hy3 into its own products, including WorkBuddy, Yuanbao, WeChat, and the game assistant for Path of Exile: Advent, and platform support for OpenRouter and Clin

Tencent’s Hunyuan team released the full Hy3 model on July 6, 2026, under the permissive Apache 2.0 license, reversing the more restrictive terms associated with its April preview. The model uses a 295-billion-parameter Mixture-of-Experts architecture with 21 billion active parameters per forward pass, top-8 routing ac The release followed roughly ten weeks of feedback and post-training changes informed by more than 50 Tencent product teams, while the underlying architecture reportedly remained unchanged. VentureBeat frames the practical impact around production deployment economics and reliability, but notes that benchmark results w

Third-party leaderboard llm-stats.com ranks Hy3 at position 46 overall with a composite LLM Stats Score of 43.0 and a blended price of $0.16 per million tokens. Capability-tier placement shows it in the top half for Search (12 of 90) and Long Context (14 of 119), while sitting mid-pack for Math (33 of 327), Coding (38 The page also surfaces specific Hy3 benchmark results sourced from its Hugging Face card, including USAMO 2026 (30.24/42, rank 3), DeepSearchQA (0.91/1, rank 5), GPQA (0.90/1, rank 25), and IMO-AnswerBench (0.90/1, rank 3), with a note that scores are self-reported by the model provider. Cost-efficiency positioning com

Tencent's Hy3 (Hunyuan 3) is a 295B-parameter MoE flagship that activates 21B parameters per token via 192 routed experts under top-8 routing plus an always-active shared expert, paired with a 256K-token context window and three selectable reasoning-effort modes. The full release ships under Apache 2.0, removing the territorial restrictions that applied to the earlier April preview, and includes a dedicated Multi-Token-Prediction submodule of roughly 3.8B parameters for inference-time acceleration. These details are drawn from the vendor model card as summarized by a third-party model-timeline profile. Vendor-reported benchmarks for Hy3 include GPQA Diamond at 90.4%, SWE-bench Verified at 78.0% (with about 4% scaffolding variance across CodeBuddy, Cline, and KiloCode), and SWE-bench Pro at 57.9%, while Humanity's Last Exam is reported only with tools at 53.2 and no independent no-tools figure is yet available. Several commonly tracked metrics, including MMLU-Pro, Terminal-Bench 2.x, and MMMU-Pro, are not reported by the vendor for Hy3, and the aggregator cautions that a quoted MMLU-Pro of 65.8 pertains to the 5-shot base checkpoint rather than the instruct/thinking tracked range. Estimated VRAM at Q4 quantization is around 170 GB, placing self-hosting in the multi-GPU class.

Tencent's Hy3 is a 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters, using 192 experts with top-8 routing for selective computation. It was released on July 6, 2026 according to the official OpenRouter model page, which preserves attribution to Tencent as the model creator. The model supports a 256K context window for long-horizon tasks and offers configurable reasoning effort, including a default no-think mode plus low and high chain-of-thought settings for complex math, coding, and multi-step problems. Tencent positions Hy3 for reasoning, agentic workflows, coding, document processing, financial analysis, and tool-calling that generalizes across agent scaffoldings, with an emphasis on grounded, anti-hallucination behavior.

Tencent's Hy Team released and open-sourced the Hy3 preview on April 23, 2026, following a reorganization, with a Mixture-of-Experts design totaling 295 billion parameters, 21 billion activated parameters, and a 256K-token context window. According to the Baidu encyclopedia entry, the model was deployed across Tencent The entry also records that on April 29, 2026, OpenRouter's global large-model API call-volume ranking placed Hy3 preview first overall, first in tool-calling scenarios, and second in programming scenarios, with the top five high-call-volume applications being agent- and code-related. On July 6, 2026, Tencent officiall