Tencent has open-sourced Hy3, a 295-billion-parameter mixture-of-experts model that activates only 21B parameters per token and ships under a permissive Apache 2.0 license. Released on July 6 and free to try on OpenRouter through July 21, Hy3 is aimed squarely at agentic and coding workflows. The headline is not the raw parameter count. It is that a frontier-class open model now beats a well-regarded closed rival at roughly half the size, which quietly resets what a solo builder can run, fine-tune, and ship without paying a token bill.
Background
For most of the past two years the open-weights conversation followed a familiar arc: an impressive Chinese or research-lab model would land, top a benchmark, and then trail the closed flagships on the workloads that actually pay the bills, namely long agent runs and real codebase edits. Hy3 lands in a different moment. Open coding models have been compounding fast, from LongCat-2.0's trillion-parameter push to Zhipu's GLM-5.2 with its million-token context, and the gap to closed models has narrowed to the point where the differentiator is efficiency, not just capability.
Hy3 is Tencent's Hunyuan team leaning into that shift. Rather than chase the largest possible dense model, it uses a sparse mixture-of-experts design that keeps inference cheap while the total parameter pool stays large. The weights are on Hugging Face, the model card and code are on GitHub, and the Apache 2.0 license means commercial use, modification, and redistribution are all explicitly permitted. That combination, frontier results plus a genuinely permissive license, is still rare.
Deep Analysis
Sparse activation is the whole trick
The number that matters most on Hy3's spec sheet is not 295B, it is 21B. That is how many parameters actually fire for any given token, thanks to the mixture-of-experts routing that selects a small subset of the network per step. A 3.8B multi-token-prediction layer sits on top to speed up generation. The practical effect is that Hy3 reasons with the breadth of a very large model but costs closer to a 21B model to serve. This is the same architectural bet behind JetBrains Mellum2 and a growing list of open MoE releases, and it is why "half the size" comparisons keep showing up. VentureBeat found Hy3 beats GLM-5.2 on most tasks at about half the active footprint.
The benchmarks, and where the asterisks are
Hy3 posts 78% on SWE-Bench Verified, the benchmark that measures whether a model can resolve real GitHub issues, and 90.4% on GPQA Diamond for graduate-level reasoning. In blind expert evaluation it scored 2.67 out of 4, edging out other recent open-weights coding models. Just as important for anyone running long autonomous sessions, its measured hallucination rate dropped from 12.5% to 5.4% versus the prior generation, which translates directly into fewer wrong-answer retries mid-run.
The asterisk is coding depth. The same VentureBeat testing that found Hy3 winning "everywhere except coding" is a signal worth taking seriously: Hy3 is a superb generalist agent model that also codes well, not necessarily the single best pure-coding model you can download. Developer Simon Willison's hands-on notes land in the same place, praising its agentic behavior while flagging that dedicated coding specialists still have an edge on the hardest programming tasks.
256K context changes what agents can hold
Hy3's 256K-token context window is large enough to keep an entire mid-sized codebase, a multi-step plan, and a running log of tool outputs in memory at once. For agent builders this is the difference between an assistant that constantly re-reads files and one that keeps the whole task in working memory. It is not the million-token frontier that GLM-5.2 reaches, but for the majority of real agent runs, 256K is the point where context stops being the bottleneck and model reasoning takes over.
Free for two weeks is a distribution strategy
The free OpenRouter window through July 21 is not just generosity, it is how an open model buys mindshare. Point any OpenAI-compatible client at the free Hy3 route, set the model string to tencent/hy3, and it drops into an existing agent stack with no code changes. That frictionless trial path, combined with downloadable weights for anyone who wants to self-host later, is the modern open-weights playbook: earn adoption with a hosted free tier, then convert the serious users to self-hosting or fine-tuning once they trust the model.
Impact on Creators
If you build apps, agents, or automations, Hy3 lowers the floor on what independent creation costs. An Apache 2.0 model that clears 78% on SWE-Bench Verified means the reasoning core of a coding agent, a research assistant, or a multi-step content pipeline can now run on infrastructure you control, with no per-token fee and no license that claws back commercial rights. For anyone who has watched a closed-model bill climb as an agent loops, that is the headline.
The practical move this week is a bake-off. Route Hy3 alongside whatever model currently powers your workflow, run both against a real task you already understand, and compare the outputs, the retry count, and the total tokens burned. Because the OpenRouter trial is free until July 21, the only cost is your time, and the results will tell you whether Hy3 is ready to take over a production job or belongs in the "watch closely" pile. Given its generalist strength, the workflows most likely to switch are multi-tool agents and long research or document tasks, rather than the hardest pure-code refactors where specialists like Kimi K2.7-Code still lead.
Key Takeaways
1. Hy3 is a 295B mixture-of-experts model that activates only 21B parameters per token, so it reasons big but serves cheap.
2. Apache 2.0 licensing plus downloadable weights make it fully usable for commercial products, self-hosting, and fine-tuning.
3. It scores 78% on SWE-Bench Verified and beats GLM-5.2 at roughly half the active size, but dedicated coding models still edge it on the hardest programming tasks.
4. A free OpenRouter window through July 21 makes a same-week bake-off against your current model essentially costless.
What to Watch
The near-term signal is what happens after July 21, when the free route ends and creators decide whether Hy3 is worth self-hosting or paying to run. Sustained adoption after the trial closes is the real vote of confidence, not the launch-week benchmark chatter. Watch too for the fine-tunes: an Apache 2.0 frontier model with strong agentic behavior is exactly the kind of base that the community turns into specialized coding, writing, and domain variants within weeks, the same pattern that followed earlier open releases.
The larger trend Hy3 confirms is that the open-weights frontier is now competing on efficiency, not just raw capability. When a downloadable model beats a respected closed rival at half the size, the question for builders stops being "can open models keep up" and becomes "which open model fits my workload." That is a healthier question, and it is the one creators will be asking for the rest of 2026.
