Tencent Hy4 preview: 770B open weights, Apache 2.0

The reason to care about open weights was never that you were going to run them in your garage. It was leverage. Tencent just released Hy4 preview — a 770B-parameter mixture-of-experts flagship under Apache 2.0, weights on Hugging Face, hosted API live the same day at under a dollar per million input tokens. That combination is what makes a fallback real instead of theoretical.

What actually happened

Per Tencent's model card, Hy4 preview is an MoE model with 770B total parameters and 49B activated per token. The backbone runs 78 layers: the first is a dense FFN, the remaining 77 are MoE with 256 routed experts plus one shared expert, top-8 routed per token. Context length is 1M. There is a native MTP layer (10B total, 0.7B activated) for speculative decoding, and an FP8 quantized variant shipped alongside the full-precision weights. Deployment targets are vLLM and SGLang with an OpenAI-compatible API.

The architecture borrows openly: Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, and identity Hyper-Connections on the residual path.

On price, OpenRouter lists it at $0.834 per million input tokens, $2.501 per million output, $0.042 per million on cache reads, with a 1,048,576-token window and 64K max completion. One provider — Tencent Cloud.

Tencent's headline evidence is an internal blind evaluation: 163 in-house experts rating 203 engineering tasks, where Hy4 preview scored 2.99 against GLM 5.3's 2.92 (46.8% wins, 12.8% ties, 40.4% losses) and Kimi K3's 2.94. That is a vendor grading its own model with its own staff. It is a reason to run your own eval, not a result.

Tencent also publishes its known problems, which is unusual and worth noting: the model spends longer than necessary reasoning through complex tasks, and over-verifies its own work. Both cost you tokens.

Why open weights matter for your business

You are almost certainly not going to self-host 770B parameters. 49B active per token still means multiple high-memory GPUs, and that's before you staff it. So the practical value here isn't sovereignty — it's price discipline and an exit that actually exists.

An Apache 2.0 license means nobody can revoke your right to run it. If Tencent Cloud raises prices or gets export-controlled out of your reach, another host can serve the same weights, and the weights you mirrored today still work. Compare that to a closed API, where deprecation is a blog post and your only move is a migration.

The move is boring and specific: pick one real workload — document extraction, ticket triage, a code review pass — and run it against Hy4 preview through an OpenAI-compatible client this week. Log cost per completed task, not cost per token. If a frontier-adjacent open model at $0.83/M gets you 90% of the result at a fifth of the spend, that's your negotiating position with every other vendor on the list.

Sources: Tencent Hy4-preview model card, OpenRouter.