Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context
What happened
Tencent’s Hy team released Hy3, a massive 295 billion parameter Mixture-of-Experts (MoE) language model. Although the full model size is 295 billion parameters, each token activates just 21 billion parameters to keep computation efficient. Hy3 supports an unusually long 256,000 token context window and is available under the permissive Apache 2.0 license. It targets tasks that demand complex reasoning, agency, and handling extended input sequences. Early benchmarks report a 78.0 score on SWE-Bench Verified with lower hallucination rates. The model can be tested freely on OpenRouter until July 21, 2026.
Why it matters
Mixing MoE architecture with a giant parameter count while activating only a fraction per token trims inference costs for large models. This lets Hy3 handle long documents and complex reasoning without the usual computational explosion of full 295 billion-parameter models. The massive 256K context window is rare and practical for workflows that require extended memory, such as detailed legal review, technical documentation processing, or multi-turn dialogue systems. Tencent releasing the model under Apache 2.0 also pressures commercial LLM providers by opening access to high-scale infrastructure for builders experimenting at long context length and agentic tasks. Lower hallucination rates mean it could raise expectations for more reliable, trustworthy outputs in complex use cases.
What to watch next
Monitor user feedback and adoption on OpenRouter to evaluate real-world performance and efficiency at scale. Watch whether other AI vendors accelerate support for large context windows and MoE architectures to keep pace. Performance data on reasoning and hallucination reduction will determine if Hy3 shifts standards for trust and accuracy in public large-scale models. Tencent’s ongoing updates could also influence competitive pricing and licensing strategies in the long-context LLM space. Operators building workflows that need consistent, large-context inference should consider testing Hy3 as an alternative to monolithic transformer models.
AI Quick Briefs Editorial Desk
