Hy3

2026-07-06 · Publisher : Tencent · Domain : Artificial Intelligence

Tencent releases Hy3, a mixture-of-experts model with 295 billion parameters, 21 billion of them active, under the Apache 2.0 license, with a 256,000-token context, designed for agents and production use.

### Context Hy3 is the final version of the Hy3 Preview launched in late April 2026, tuned from feedback from more than 50 Tencent products. The model also marks a license change: the community license of Hunyuan-A13B (a year earlier) excluded the European Union, the United Kingdom and South Korea, whereas Apache 2.0 has no geographic restriction. Against the closed references of summer 2026 (GPT-5.5, Claude Opus 4.7 and 4.8, Gemini 3.1 Pro according to other vendors' tables published in the same period), the vendor mostly compares itself to open models. ### What the model brings - Production reliability: fixes for basic defects in tool calls and output formats; on SWE-Bench Verified, the accuracy gap between agent frameworks (CodeBuddy, Cline, KiloCode) stays within 4%. - Fewer hallucinations according to the vendor: the rate drops from 12.5% to 5.4% and common-sense errors from 25.4% to 12.7% on internal evaluations of real-world scenarios. - Better conversation tracking: the problem rate on internal multi-turn tests drops from 17.4% to 7.9%, with gains on MRCR for long dialogues. - Blind evaluation by 270 experts on work tasks: 2.67 out of 4 for Hy3 versus 2.51 for GLM-5.1 (vendor test, open model compared). - Adoption: Tencent states that Hy3 reached first place in OpenRouter's global usage ranking during the week of launch (Tencent press release, not verified). ### Architecture and specifications - Architecture: MoE with 295 billion parameters, 21 billion active, plus 3.8 billion MTP layer parameters; 80 layers, 192 experts of which 8 are activated per token, grouped-query attention (64 heads, 8 key-value heads). - Context: 256,000 tokens. Modalities: text. Published precision: BF16, with an Hy3-FP8 version. - License: Apache 2.0 (free commercial use, no user threshold). - Weights: Hy3 and Hy3-FP8 on Hugging Face, ModelScope, GitCode and CNB; about 598 GB in BF16 and 300 GB in FP8 according to Simon Willison. Knowledge cutoff: not disclosed. ### Cost and access According to a research summary, Tencent's API is priced at 1 yuan per million input tokens, 4 yuan for output and 0.25 yuan for cached input (secondary source). Tencent's press release cites 0.1288 dollars for input and 0.5336 dollars for output per million tokens for worldwide distribution. | Item | Price | | --- | --- | | Input (Tencent API, yuan, secondary source) | 1 per million tokens | | Output (Tencent API, yuan, secondary source) | 4 per million tokens | | Input / output (USD, Tencent press release) | 0.1288 / 0.5336 per million tokens | | BF16 / FP8 weights | about 598 GB / 300 GB | ### Benchmarks | Indicator published by the vendor | Before | Hy3 | | --- | --- | --- | | Hallucination rate (internal evaluation) | 12.5% | 5.4% | | Common-sense errors (internal evaluation) | 25.4% | 12.7% | | Problems in multi-turn conversation (internal test) | 17.4% | 7.9% | | Blind evaluation, 270 experts (out of 4) | GLM-5.1: 2.51 | 2.67 | Source: Hy3 Hugging Face model card, Tencent internal measurements. The model card's comparison charts are images: no numerical table comparing Hy3 to closed models is available in text. ### Versus closed models There is no numerical table in text comparing Hy3 to a closed model in the sources consulted: the vendor's comparison is made in images and against open models (GLM-5.1). Tencent writes that Hy3 outperforms models of comparable size and rivals flagship open models with 2 to 5 times more parameters; this is a vendor claim. The numerical progress concerns reliability (hallucinations, tool formats, conversation tracking) more than raw capability. On cost, the output price of 4 yuan per million tokens (secondary source) places Hy3 in the category of very cheap models. For the enterprise, the main appeal is the combination of Apache 2.0, 21 billion active parameters and a 256,000-token context, which allows internal deployment without geographic restriction. ### Key takeaways Hy3 is Tencent's first flagship open model under Apache 2.0, oriented toward production reliability. Public evidence of its level against closed models is lacking: it is in images or in internal tests, and should be confirmed by independent rankings.

Tags : LLM, Open Source, Tencent, Hunyuan, Hy3, Apache 2.0, MoE, Agents