News · 2026-07-09

Tencent open-sources Hy3, a lean mixture-of-experts model that punches above its weight

Tencent has open-sourced Hy3, a large language model built on a mixture-of-experts design with 295 billion total parameters but only 21 billion active for any given token, released under the permissive Apache 2.0 license. Tencent says it competes with models five times its size while supporting a 256K-token context window, and it is already integrated into the company's WorkBuddy, Yuanbao, and Marvis products.

Key facts

• What: Hy3, an Apache 2.0-licensed mixture-of-experts model from Tencent.
• Size: 295 billion total parameters, 21 billion active per token, 256K context.
• Claim: Competitive with models roughly five times its size.
• Primary source: Tencent's release, discussed on Hacker News (front-page #12).

The architecture is the point. A traditional "dense" model runs every parameter for every token, so a 300-billion-parameter model costs 300 billion parameters' worth of computation on each word. A mixture-of-experts model instead splits its parameters into many specialized "experts" and a router that, for each token, activates only a few of them. Hy3 holds 295 billion parameters' worth of knowledge but fires only 21 billion at a time -- roughly a fourteenth of the total. The analogy is a large hospital with dozens of specialists on staff: any single patient sees only the two or three relevant doctors, not all of them, so the institution is knowledgeable without every visit costing the whole payroll.

That is why the "competitive with models 5x its size" framing is plausible rather than marketing bluster: Hy3 gets the breadth of a very large model with the serving cost of a much smaller one. Combined with a 256K context window -- enough to hold a substantial codebase or a book in working memory at once -- it is designed to be cheap to run at scale, which matters for the products Tencent has already wired it into.

The genuinely industry-relevant detail is the license. Apache 2.0 permits commercial use, modification, and redistribution with minimal strings attached, unlike the more restrictive licenses some open-weight models ship under. That puts Hy3 in the same permissive tier as the strongest Western open releases and continues a steady drumbeat of capable, openly-licensed models coming out of Chinese labs -- a trend that keeps pressure on closed frontier providers' pricing. See our lesson on open-weight models for why licensing terms, not just benchmark scores, decide whether a model gets adopted.

Why it matters: the practical frontier for most companies is not the single most capable model but the best model they can run affordably and legally on their own terms. An openly-licensed MoE that claims near-frontier quality at a fraction of the serving cost is exactly the kind of release that erodes the moat around paid APIs. The honest caveat: the "5x its size" claim came from Tencent's own materials, and independent head-to-head benchmarks were thin at launch -- the HN discussion focused more on Tencent as an AI company than on verified internals. As always with a fresh release, the vendor's numbers are a starting point for scrutiny, not a verdict.

Primary source, verified: read the paper →

Ground Truth. (2026, July 9). Tencent open-sources Hy3, a lean mixture-of-experts model that punches above its weight. Ground Truth. https://groundtruth.day/news/tencent-hy3-open-moe-punches-above-weight.html

@misc{groundtruth:tencent-hy3-open-moe-punches-above-weight,
  title  = {Tencent open-sources Hy3, a lean mixture-of-experts model that punches above its weight},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/tencent-hy3-open-moe-punches-above-weight.html}
}

Topics: Tencent · open-weights · mixture-of-experts · model-release · China