Tencent open-sources Hy4 preview: 770B MoE, 1M context, Apache 2.0

Tencent's Hy team has put out Hy4 preview, a 770B-parameter mixture-of-experts model that activates 49B per token and handles a 1M-token window, with weights on Hugging Face under Apache 2.0 and paid API access through Tencent Cloud and OpenRouter.

Key points

• Model IDs: tencent/Hy4-preview and tencent/Hy4-preview-FP8 on Hugging Face (licence field apache-2.0); the API slug is tencent/hy4-preview on OpenRouter, and Tencent Cloud TokenHub serves it too.
• Architecture per the model card: 770B total, 49B active; 78 layers (first dense, 77 MoE) with 256 routed experts and one shared expert, top-8 routing per token, plus a built-in 10B-parameter MTP layer for speculative decoding.
• Context: 1,048,576 tokens (the vLLM recipe lists it as 1024K). Tencent's API pricing: $0.834 per 1M input, $2.501 per 1M output, $0.042 per 1M on cache hits; users of WorkBuddy and CodeBuddy get a fortnight of free use from launch.
• Vendor's blind eval (163 Tencent staff scoring 203 engineering tasks): 2.99 out of 4 against GLM-5.3 at 2.92 and Kimi K3 at 2.94, with a 46.8% win rate vs GLM-5.3; the card also self-reports GPQA Diamond 92.3 and Terminal-Bench 2.1 at 85.4.
• Vs Hy3 preview (295B total, 21B active, 192 experts, 256K context, Tencent's own community licence): about 2.6x the parameters, 4x the context, 256 experts, and a switch to plain Apache 2.0. Tencent says it scaled training data too, and that the model found and fixed bottlenecks in its own inference system for a claimed 31.8% end-to-end throughput gain.
• Local: vLLM 0.29+ and SGLang both have prebuilt images, but the vLLM recipe targets 16x B200 or 8x B300 at tensor-parallel 8, so outside a datacenter this is a rent-it model.

“we would rather ship early and hear what breaks.” — Tencent Hy team, Hy4-preview model card on Hugging Face

Why it matters

For a small studio the win is not running 770B on your own hardware, it is that a frontier-class agentic model now ships under plain Apache 2.0 rather than the custom community licence Hy3 preview carried, so any host can serve it and Tencent's own $0.834/$2.501 rate becomes a ceiling other providers can undercut. A million-token window at that price makes whole-repo context cheap enough to try on real client work, and 49B active out of 770B keeps per-token compute closer to a mid-size dense model. Treat the 2.99-vs-2.92 edge over GLM-5.3 as Tencent's own claim from its in-house raters until independent harnesses weigh in.

This is an AI-written summary of the reporting credited above, read and edited by Nicholas before publishing. The facts and any quote belong to the original; the wording is ours. Read the original.