Tencent releases Hy3, a 295B-parameter MoE model with 21B active parameters and 256K context
The Apache 2.0–licensed model outperforms similar-size models and rivals larger open-source models, with a 598GB full-size and 300GB FP8 quantized release on Hugging Face.
1 source · cross-referenced
- Tencent’s Hy3 is a 295B-parameter Mixture-of-Experts model with 21B active parameters and 3.8B MTP layer parameters.
- It outperforms similar-size models and rivals flagship open-source models with 2–5× parameters.
- The full model is 598GB on Hugging Face; the FP8 quantized version is 300GB.
- Context length is 256K tokens.
- Available for free on OpenRouter until July 21, 2026.
Tencent’s Tencent Hy Team has released Hy3 under the Apache 2.0 license. The model is a Mixture-of-Experts (MoE) architecture with 295B total parameters, 21B active parameters, and 3.8B parameters in the MTP layer.
According to the announcement, Hy3 was developed following a preview launch in late April, during which the team collected feedback from more than 50 internal products and scaled post-training with higher-quality data.
The developer documentation reports that Hy3 outperforms similar-size models and rivals flagship open-source models whose total parameters are two to five times larger. It also shows significant gains in utility across various internal products and productivity tasks.
The full-size model is 598GB on Hugging Face, while the FP8 quantized version is 300GB. The model supports a context length of 256K tokens.
Hy3 is available for free on OpenRouter until July 21, 2026, where users can test it via chat interfaces.
- Jul 7, 2026 · arXiv cs.CL
Gemma 4 released with multimodal, encoder-free architecture and reasoning traces
Trust79 - Jul 5, 2026 · Simon Willison’s Weblog
Newer Anthropic models increasingly emit malformed tool calls for third-party edit schemas
Trust79 - Jul 2, 2026 · Simon Willison — everything
Claude Sonnet 5 launches with 1M token context window, new tokenizer, and adaptive thinking enabled by default
Trust79