Tencent's Hy4: a 770B open model that drives your desktop

Hunyuan's flagship adds command-line and GUI control and lands near Claude on OSWorld, callable on OpenRouter today; plus Claude's API stops resending chats.

Nowline NOV 4 7:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Hy4 can drive your desktop now

    Tencent's new Hunyuan flagship does computer use for the first time, operating both the command line and on-screen apps, with extra training on Blender and CAD for long, multi-app jobs. You can hand it a GUI workflow you'd otherwise have scripted by hand and let it click through.

  • 770B open weights, days away

    Hy4 is a 770B-parameter mixture-of-experts (~49B active) with a 1M-token context. The preview weights are already on Hugging Face under Apache 2.0, and Tencent says the full Hy4 weights land “in the coming days” with vLLM, SGLang and Inferact support — so self-hosting is imminent, not theoretical.

  • Call it today, but know the ceiling

    Hy4 is live on OpenRouter now (~$0.42/M input, $1.25/M output, 1M context, image input), and Tencent's CodeBuddy is free Nov 4–15. On its own numbers Hy4 hits 78.9 on OSWorld 2.1 — ahead of GPT-6.1 Sol's 75.2, just under Claude Opus 5.5's 81.8 — but trails on coding, where Kimi K3.1 Pro's 66.8 on Terminal Bench beats Hy4's 44.4.

  • Claude's API stops making you resend the chat

    Anthropic's new Message Threads beta (Messages API) lets you send a conversation once, then pass only the new turns plus a thread ID — cutting a long agent request from ~800KB to ~1.8KB and dropping median time-to-first-token ~35% on 500K–1M-token calls. Billing is unchanged and threads expire after ~5 minutes idle, so keep your own copy to recover on thread_not_found.