Tencent open-sources 295B MoE model Hy3, only 21B active params, 256K context
Decision Brief
What changedTencent Hy team open-sources 295B MoE model Hy3, activating only 21B params per token with a 256K context window.
Why it mattersHy3 scores 78.0 on SWE-Bench Verified, Apache 2.0 licensed, ideal for low-cost local inference or agent tasks.
Who should careOpen-source model users
Affected stackNo specific stack identified
Source confidenceMedium · Reliable media or first-hand reporting
Tencent Hy team launched Hy3, a 295B MoE model activating only 21B params per inference, reducing compute. Apache 2.0, 256K token context, optimized for reasoning, agents, and long texts. Achieves 78.0 on SWE-Bench Verified, claims lower hallucination. Free trial on OpenRouter until July 21, 2026. For teams using open-source models, Hy3's 21B effective params means latency and memory near medium models but knowledge capacity of a 300B model. 256K context suits code or long docs. API devs can test free via OpenRouter; private deployment gets flexible commercial rights under Apache 2.0.
Summary basis: official / RSS sourceCompiled from the source scope noted above; the original remains authoritative.
Sources
- MarkTechPost
Fast research-paper and ML tooling summaries, useful for infra and agent updates.
- MarkTechPost