Tencent's Hunyuan team dropped a new flagship model this week called Hy4 preview, and the spec sheet is the kind that makes you re-read it twice: 770 billion total parameters, a 1 million token context window, and — the part that actually matters for anyone outside a research lab — it's open-sourced.
It's already live on Tencent Cloud's TokenHub and on OpenRouter, so you don't need a Tencent account or a data center to try it.
The Numbers, Translated
770B total parameters sounds enormous, but the number that matters more is the activation parameter count: 49B. Hy4 preview is almost certainly a mixture-of-experts (MoE) model, meaning it doesn't run all 770B parameters for every token — it routes each request through a much smaller active slice, roughly 49B worth. That's the trick behind how a model this large can be usable at all outside a supercomputer: you get frontier-scale knowledge with mid-size inference cost.
The 1 million token context window is the other headline number, and it puts Hy4 in the same conversation as the handful of models that can genuinely hold an entire codebase, a full legal contract set, or a season of meeting transcripts in a single prompt without chunking.
Why Everyone's Racing Toward Bigger Context Windows
Context length has quietly become one of the most competitive metrics in AI right now, alongside benchmark scores. The reason is practical, not academic: a bigger context window means less engineering work stitching together retrieval systems, less risk of a model forgetting something you told it three messages ago, and fewer surprises when a task turns out to need more background than you expected.
For a coding agent specifically, this matters a lot. A 1M-token window can hold a genuinely large repository's worth of source files at once — not just the function you're editing, but the surrounding modules, the tests, and the docs that explain why the code is shaped the way it is.
Open Source Is the Real Story Here
Plenty of labs ship huge models. Fewer open-source them. Tencent releasing the weights for a 770B model puts real pressure on the rest of the frontier-model landscape — it's now one more capable option that developers, researchers, and companies can self-host, fine-tune, or inspect line by line instead of trusting a black box behind an API.
That matters for anyone building agent infrastructure. Open weights mean you're not locked into one vendor's pricing, one vendor's rate limits, or one vendor's decision to deprecate a model version out from under you.
The Bigger Picture
A year ago, a 1M-token open-source model would have been a research curiosity. Now it's a Tuesday release that shows up on OpenRouter within days. The pace of open-weight releases keeps compressing, and each one lowers the floor for what "good enough to build on" looks like.
If you want to see what that capability looks like when it's put to work — not just benchmarked, but actually running tasks end to end — that's what AI agents are for.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.