Tencent Releases Hy4 Preview, an Open Weight 770B Model

Hy4 Preview weights take up 1.56 TB on Hugging Face. That figure sums up what Tencent presented on 29 August better than any benchmark table: an open weight language model with 770 billion total parameters, 49 billion active per token and a one million token context window. Text only, no vision.

Simon Willison covered the release the same day on his weblog and added the detail that rarely makes it into the official announcement: how much the model has grown compared with the previous generation.

From July to August, 2.6 times bigger

Hy3 arrived in July with 295 billion total parameters, 21 billion active, 256,000 tokens of context and 598 GB of weights (model card on Hugging Face). Hy4 Preview multiplies total size by 2.6, more than doubles active parameters and quadruples the context window, according to the model card and the Tencent announcement.

The ratio between total and active parameters (49B out of 770B, around 6%) points to a very sparse mixture of experts, in line with Hy3 (21B out of 295B, 7%). The change is one of scale, not of approach: Tencent has stretched the same recipe a month and a half later.

The 1.56 TB figure says something too. It works out at roughly two bytes per parameter, consistent with weights published in 16 bits without quantisation. An 8 bit or 4 bit quantised version would cut the memory bill considerably, but serving the model is still multi node GPU territory, not workstation territory.

The chat template as a spec sheet

Willison says he has started reading model chat templates to understand what they can do before testing them. In Hy4 Preview, the chat_template.jinja file includes a conditional block that checks whether the reasoning_effort variable is defined and assigns a default value when the client does not send it.

It looks like a minor detail and it is one of the most useful things you can inspect in a new model. The template describes which roles it accepts, how it expects tool definitions to arrive and which control parameters it understands. That information shapes the real integration work long before any benchmark score does, and it is usually published even when the documentation lags behind.

Who it is useful for

Two things get mixed up when people talk about open weights. Downloading 1.56 TB and holding 770 billion parameters in memory is out of reach for a small team. What a weights release does allow is for an inference provider, a lab or a company with its own infrastructure to serve it without depending on the Tencent API, audit it and tune it to a specific domain. For everyone else, the realistic route is waiting for it to show up at some provider via API.

The one million token window is the other headline, and this is where the enthusiasm should come down a notch. Claude Opus 4.8 already offers an optional 1M window, so the number alone no longer sets anyone apart. What does set a model apart is what it does with that window: whether it keeps its accuracy retrieving details buried in the middle of the context or degrades past a certain point. That data is not in the announcement and will have to be checked against independent evaluations.

Our take

The striking part is not the size, it is the pace. Going from 295B to 770B in little more than a month suggests the open weights release cycle is moving faster than the ability of most teams to evaluate it properly. We keep building product on Claude, because the surrounding tooling ecosystem (MCP, skills, subagents) matters more day to day than a parameter count. Even so, an open model at this scale with a 1M context window is exactly the kind of piece worth keeping an eye on for workloads where the data cannot leave the building.