The takeaway: Tencent released Hunyuan Hy3 this week — an open-weights, Apache-licensed mixture-of-experts model (295B parameters, 21B active, 256K context) whose vendor-claimed benchmark scores would put it within reach of models many times its running cost. The claims are Tencent's own, and no independent evaluation has been published yet. But the release itself is the story: in the same week Alibaba's Claude Code ban took effect and Beijing's agent rules bit, China's other tech giant answered the frontier-pricing war with the opposite move — giving the weights away.

What Tencent shipped

Hy3's research page went up on Monday, with the weights released under Apache 2.0 — free for commercial use. The architecture is the now-familiar sparse mixture-of-experts bet: 295 billion total parameters with only 21 billion active per token, which is what lets a model of this claimed capability run at small-model speeds and costs. Context window: 256K tokens.

The headline numbers — all Tencent's own, and worth repeating that until independent evaluators publish, vendor tables are marketing with error bars — include 57.9 on SWE-Bench Pro (up from 46.0 for its preview build), 78.0 on SWE-Bench Verified, 84.2 on BrowseComp and 90.4 on GPQA Diamond. Tencent's pitch is not "we beat the frontier" but the sharper commercial claim: a smaller open model that gets close enough, for a fraction of the cost. The developer crowd took notice — the release front-paged Hacker News at over 450 points on Thursday.

One comparison we'd caution against making too quickly: putting Tencent's 57.9 SWE-Bench Pro next to the numbers in OpenAI's launch tables this week. Different vendors run these benchmarks under different harnesses and settings; cross-vendor comparisons only mean something when one evaluator runs everything the same way. Hy3 has not yet appeared on Artificial Analysis's leaderboards — when it does, we'll report where it actually lands.

Why the timing reads as a statement

Hy3 landed in the middle of Alibaba's worst week: the Claude Code employee ban taking effect, the Qwen agent shutdown under Beijing's new rules, and the distillation feud with Anthropic still open. Against that backdrop, Tencent's move continues the pattern we examined in the margin-collapse argument: Chinese labs competing not by matching Western frontier pricing but by undercutting it — or, in the open-weights case, zeroing it. In one week the frontier price card has been rewritten — $10/$50, $5/$30, $2/$6 in the API tier — and the open-weights floor keeps pressing from below. A capable Apache-licensed model doesn't need to beat GPT-5.6 or Fable 5 to matter; it needs to be good enough to make "free, run it yourself" a serious answer for another slice of the market.

The independent numbers will tell us whether Hy3 clears that bar. Until then: a real release, a real licence, and claims that deserve exactly as much confidence as every other vendor table this week earned — which is to say, verification first.

Disclosure: On The Wire runs on Anthropic models. We flag it when we cover the companies competing with them.