Here’s my real verdict. This is my honest Tencent HY3 review after testing it for hours. HY3 is a brand-new open-source model from Tencent (China), and right now you can use it free via OpenRouter. Is it any good? Yes and no — here’s the real story.
Short version: it’s great for free agentic tasks, but it’s not a frontier model for polished builds.
Last updated: July 2026.
Key takeaways
- HY3 is a 295B mixture-of-experts model (21B active), open source, and designed for agent tasks.
- It’s free via OpenRouter (until July 21), plus in Hermes agent and Kilo Code.
- Weak spot: coding and game builds look basic and glitchy next to GLM 5.2, Fable 5 and Opus 4.8.
- Strong spot: agentic tasks — browsing, learning skills, running tools, and the HY3 Coder workflow.
What Is Tencent HY3?
HY3 is Tencent’s new open-source model out of China. Technically it’s a 295-billion-parameter mixture-of-experts model with around 21 billion active parameters, and the model details make clear it’s built for agent capabilities rather than pure frontier coding.
The headline is that it’s open source and free to try — which, whatever its limits, makes it worth a look.
How to Use HY3 for Free
There are a few free routes:
- OpenRouter — search “HY3”, grab the free API and plug it in (free until July 21)
- Hermes agent — available in News Research for free
- Kilo Code — also free
For more free options, see my best free AI model for Hermes guide.
Tencent HY3 Review: The Honest Verdict
I ran HY3 through my Goldie Bench tests, coding and wrestling with it for three to four hours. Here’s the honest take: it’s awesome that it’s free and open source, but is it one of the best models I’ve used? No. Is it right up there with the frontier? Not really.
Where it disappoints is quality. The landing page it built wasn’t bad, but the game and 3D creations felt basic and glitchy — weird lighting, no detail, not really playable. Even after a lot of back-and-forth, the outputs didn’t get to where I’d hoped.
HY3 vs GLM 5.2, Fable 5 & Fusion
Side by side on the same prompts, the gap is clear:
| Model | Build quality | Notes |
|---|---|---|
| Tencent HY3 | Basic, glitchy | Free & open source |
| GLM 5.2 | Much nicer | Better vibe, colour, detail |
| Fusion | Far more detailed | Noticeably higher quality |
| Fable 5 / Opus 4.8 | Frontier | Not a fair fight — HY3 is open source |
For daily building of my Agent OS, I’d reach for GLM 5.2, Claude, Fable 5 or Opus 4.8 every time. See my Kimi K3 vs Fable 5 tests for how a stronger open-source model compares.
Where HY3 Actually Shines: Agent Tasks
Here’s the flip side — HY3 isn’t built to be a frontier coder, and once I used it agentically it was genuinely good:
- Browsing — “open up Google” worked smoothly
- Learning skills —
/learnreads pages quickly, and it’s fast - Running tools — solid at agent workflows
- HY3 Coder — a free coder built on the OpenRouter API (prompts left, preview right, saves below) — genuinely great
- Remotion videos — give it the Remotion skill and it codes full animated videos, better than a lot of open-source projects
So if you just need a free model that can browse, run tools and get through agent tasks, HY3 is pretty good.
Benchmarks vs Reality
On paper HY3 scores well on some agentic evals — MCP Atlas, BrowseComp, ClawEval (where it beats GLM 5.2). But in my own hands-on testing it didn’t feel like it lived up to those numbers on quality. That gap between benchmark and reality is exactly why I built Goldie Bench and test everything myself.
Specs & Local Running
HY3 is a 295B mixture-of-experts model with ~21B active parameters, and it’s designed for agent capabilities. It wasn’t on LM Studio yet when I tested, though local support will likely land soon — and interestingly Tencent’s older models were small (~8GB), so local running could be very doable.
Should You Use Tencent HY3?
My honest recommendation: for building serious things day to day, I’d pick GLM 5.2 first among Chinese open-source models, then Qwen. But HY3 is free, open source and solid for agentic tasks — so if you want a free model to browse, run tools and power quick agent jobs, it’s well worth testing.
I’ve built the HY3 Coder and a Hermes HY3 profile into my Agent OS inside the AI Profit Boardroom, alongside training on GLM 5.2 and other Chinese models.
New here? Start free with my AI course and community (plus 1,000+ AI agents), or grab a free strategy session.
FAQ
Is Tencent HY3 any good?
For free agentic tasks — browsing, running tools, learning skills — yes, it’s solid. For frontier-quality coding and game builds, no; it looks basic next to GLM 5.2, Fable 5 and Opus 4.8.
Is HY3 free?
Yes — it’s open source and free via OpenRouter (until July 21), plus in Hermes agent’s News Research and Kilo Code.
What is Tencent HY3?
A 295B mixture-of-experts model (about 21B active parameters) from Tencent, designed for agent capabilities rather than pure frontier coding.
HY3 vs GLM 5.2 — which is better?
GLM 5.2 produced noticeably nicer, more detailed builds in my tests. HY3 wins on being free and is fine agentically, but GLM 5.2 is the better all-round model.
Can I run HY3 locally?
Not easily yet — it wasn’t on LM Studio when I tested. Given Tencent’s older models were small, local support may come soon.
The Bottom Line
Tencent HY3 is free and capable agentically, but not a frontier model. Worth a free test; for quality builds I’d use GLM 5.2 or Claude. Get it in the Agent OS via the AI Profit Boardroom.