Tencent opened a 770B model under Apache 2.0, and the weights alone are 1.56TB

The weights of Hy4 preview went up on Hugging Face on August 27 and the model card was completed the next day. It is the largest open-weights release the Tencent Hy team has made, and the two numbers that describe it answer different questions.

What actually has to fit

The download is 1.56TB, spread across 131 safetensors files in BF16. The FP8 build is 0.81TB across 130 files. Neither of those is a laptop, a Mac Studio or a single accelerator, and no quantization on the model card brings it into that range.

So read this release as an open model you host, not one you run. The distinction matters because open weights are usually reported as though they were a home option, and at this size the practical choice is between renting the hardware and renting the model from somebody who already has.

770B total, 49B active

The backbone is 78 layers. The first uses a standard dense feed-forward network and the remaining 77 replace it with a mixture of experts, each holding 256 routed experts and 1 shared expert. Every token activates the top-8 routed experts plus the shared one, which is where the 49B figure comes from.

Alongside the backbone sits one native multi-token-prediction layer, 10B parameters of which 0.7B activate, built in for speculative decoding. Attention is Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, and the residual pathway uses identity Hyper-Connections across four residual streams.

• Context length 1M tokens, hidden size 6144, 64 attention heads.
• Vocabulary of 120,832 tokens.
• Indexer top-k of 2048, with query and key-value compression at 2048 and 512.
• Deployment documented for vLLM and SGLang, through Docker containers.

The license is the part that travels

Hy4 preview and Hy4 preview-FP8 are published under Apache 2.0, on Hugging Face, ModelScope, GitCode and CNB. Apache 2.0 carries no field-of-use restriction and no separate commercial license to negotiate, which is not the case for every model that calls itself open.

At this parameter count that combination is rare. What stands between a company and using this model is a hardware budget, not a legal review, and those two obstacles are not the same kind of problem.

Tencent measured it against Chinese models, and measured it itself

The card reports a blind side-by-side evaluation run inside Tencent: 163 internal experts rated model outputs on 203 engineering tasks. Hy4 preview came out at 2.99 on average against GLM 5.3 at 2.92, with 46.8% wins, 12.8% ties and 40.4% losses, and against Kimi K3 at 2.94, with 51.2% wins, 7.9% ties and 40.9% losses.

Two things follow. The comparison set is other open Chinese models rather than anything Western, which tells you where Tencent thinks the contest is. And the raters work for the company that made the model, so this is a vendor claim about its own position, not an independent result. The benchmark tables on the card are published as images, so the scores in them cannot be quoted here.

The preview label has content

Tencent names its own open problems rather than leaving the word preview to do the work. The card says the model spends longer than necessary reasoning through complex tasks and has a tendency to over-verify its own work, and that both pre-training and post-training still have real headroom.

That is worth more than a benchmark row. Over-thinking and over-verification are exactly the failure modes that turn into token bills and latency on agent work, so they are the first thing to measure on your own tasks.