Tencent open-sources Hy4: 770B MoE, 1M context, Apache 2.0
Permissive license, no MAU cap, cheap API — but ~924GB even quantized makes it a cloud play, not a local one, despite strong coding and agent scores.
Copy markdown
The weights are live on Hugging Face
Hy4 preview is a 770B-parameter MoE with ~49B active per token and a 1M-token context, published in BF16 and FP8 at tencent/Hy4-preview and mirrored on ModelScope and GitCode.
Apache 2.0 — no monthly-active-user ceiling
Unlike GLM-5.3's custom license, Hy4 ships under plain Apache 2.0 with no MAU threshold, so you can build on it and ship commercially without asking for a separate grant.
But it won't fit your workstation
It needs roughly 924GB of memory even quantized to FP8, or ~1,848GB at full precision — a multi-GPU/rack job, not a 4090 weekend. For most builders the API is the only realistic path.
Cheaper than the frontier, at $0.83/$2.50 per M
Tencent Cloud's TokenHub lists input at $0.834 per million tokens and output at $2.501, with cache hits at $0.042. It's OpenAI-compatible and also mirrored on OpenRouter.
Level with the Chinese flagships, not the frontier
Public scores land at 92.3 on GPQA Diamond and 85.4 on Terminal Bench; Tencent's own blind eval put it at 2.99 vs Kimi K3's 2.94 and GLM-5.3's 2.92 — a near-tie, and below Opus 5 and GPT-5.6 on public tests.
Elsewhere: OpenAI turns on mTLS for the API
Mutual TLS and X.509 workload identity federation are now generally available for the OpenAI API, letting services authenticate with certificates instead of long-lived API keys.