Tencent open-sources Hy4: a 770B Apache-2.0 MoE, live on OpenRouter
49B active, 1M context, no license strings, and agent scores that tie GLM-5.3 — plus Gemini's new speech-to-text and Skild's one-video robot model.
Copy markdown
Hy4 is live on OpenRouter today
Tencent's new flagship is a 770B mixture-of-experts with just 49B active per token, priced at $0.834/M input and $2.501/M output — with a 20x-cheaper $0.042/M rate for cached input. It ships a native 1M-token context and 64K max output, so you can point a coding agent at a whole repo through a single API call right now.
Apache 2.0, no strings attached
The weights are open under Apache 2.0 with no field-of-use clause and no geographic carve-out — a cleaner license than the GLM and Qwen drops builders have been squinting at all week. If you have multi-GPU hardware, you can self-host the whole thing.
Built for agents, not just chat
Hy4 posts 65.7 on SWE-Bench Pro, 85.4 on Terminal-Bench 2.1, 83.7 on MCP-Atlas, and 92.3 on GPQA Diamond — numbers aimed squarely at tool-use and coding workflows. On input price it undercuts GLM-5.3 by roughly 40%.
The catch: it's a tie, and it's heavy
Tencent's own eval scores Hy4 at 2.99/4 versus GLM-5.3 at 2.92 and Kimi K3 at 2.94 — a margin the writeup itself calls noise. It also shows over-verification tendencies and needs multi-GPU serving to run, so benchmark it on your actual workload before you switch anything over.
Elsewhere: a speech-to-text API and a one-video robot brain
Google's Gemini 3.5 Transcribe is in public preview via AI Studio — 85+ languages at about $0.005/min with speaker labels and word-level timestamps baked in (2.6% WER). Separately, Skild's S1 learns a robot task from a single video demo (66% success on unseen tasks vs 9% for language-prompted baselines), though it's early-access only for now.