Tencent HY3 Model: Honest Review + How to Use It Free (2026)
There is a new free Chinese model worth a look. The Tencent HY3 model is a brand-new open-source AI that just dropped from China — and, better still, it is free to use right now.
I have spent the last few hours building with it, testing it on GoldieBench and plugging it into my Agent OS. Here is my honest take on what HY3 is good at, where it falls short, and how to try it for free.
Last updated: July 2026.
Key takeaways
HY3 is free on OpenRouter until 21 July 2026 (also free in Hermes Agent and Kilo Code).
It is a 295B mixture-of-experts model built for agent tasks — strong on agentic workflows, basic on frontier creative work.
I have already built HY3 into my Agent OS, including an HY3 Coder workflow.
What Is the Tencent HY3 Model?
Tencent HY3 is an open-source large language model released by Tencent. It is a mixture-of-experts (MoE) model with around 295 billion total parameters and roughly 21 billion active at a time, which keeps it efficient to run for its size.
Crucially, it is designed for agent capabilities rather than pure frontier reasoning — the model details themselves point at browsing, tool use and multi-step tasks. That focus matters a lot for how you should use it.
Spec
Detail
Model
Tencent HY3 (open source)
Architecture
Mixture-of-experts (MoE)
Parameters
295B total / ~21B active
Free access
OpenRouter (free until 21 July 2026), Hermes Agent / News Research, Kilo Code
Best at
Agentic tasks — browsing, running tools, learning skills, Remotion video
Weak at
Frontier creative & game generation (basic, glitchy)
Local
Not on LM Studio yet; details still emerging
How to Use Tencent HY3 for Free
The best part is the price: nothing. Here is where you can run HY3 free today:
OpenRouter — search ‘HY3’, grab the free API and plug it in (free until 21 July 2026).
Hermes Agent / News Research — HY3 is available there for free.
Kilo Code — also free.
HY3 Coder — build a simple coder around the free OpenRouter API, prompts on the left, live preview on the right. This is one of the best free workflows with HY3.
Tencent HY3 Benchmarks vs Reality
On paper, HY3 looks strong. On agentic evaluations like MCP Atlas, BrowseComp and Claw Eval it scores well — it even beats GLM 5.2 on Claw Eval, which is an agentic workflow benchmark.
In my own hands-on testing, though, the real-world creative output does not match the benchmark hype. That gap is exactly why I built GoldieBench — to test models on real tasks rather than trusting the scoreboard.
I share the full side-by-side tests with my community, so you can see exactly how HY3 stacks up rather than guess.
What Tencent HY3 Is Actually Good At
Play to its strengths and HY3 is genuinely useful — especially since it is free:
Browsing — tell it to open a page and it handles it well.
Learning skills & running tools — fast and reliable for agent workflows.
Coding simple pages — not bad at all; I have seen worse from paid models.
Remotion video — give HY3 access to Remotion and it can code full animated videos, better than a lot of open-source projects I have tried.
Where Tencent HY3 Falls Short
Push it toward frontier creative work and the cracks show. Across the game and world-building tests I ran, the outputs felt basic and glitchy — weird lighting, low detail, not really playable or fun.
Side by side, models like Fusion, GLM 5.2, Claude, Fable 5 and Opus 4.8 produced noticeably richer, more detailed results from the same prompts. I spent 3–4 hours going back and forth and still could not get frontier-level output.
It is open source and free, so I do not expect frontier quality — but it is worth being clear-eyed about it.
Model
Type
Best for
Creative quality
Agentic tasks
Cost
Tencent HY3
Open source (295B MoE)
Free agentic tasks
Basic / glitchy
Good
Free (to 21 Jul)
GLM 5.2
Open source
Building & creative
Strong
Strong
Low / free
Fable 5
Frontier
Top creative & coding
Excellent
Excellent
Subscription / API
Claude / Opus 4.8
Frontier
Daily building
Excellent
Excellent
Subscription / API
Tencent HY3 vs Other Chinese Open-Source Models
HY3 is not the only strong open-source model coming out of China, so how does it stack up against the others?
If I had to rank the Chinese open-source models I actually reach for, I would still put GLM 5.2 first, then Qwen 3.7. DeepSeek I do not tend to use as much these days. HY3 slots in as a specialist — the one I would pick specifically when I want a free model for agent tasks rather than general building.
That is the theme with HY3: it is not trying to be the best all-rounder, it is an agent-focused model, and judged on that basis it does well.
My Verdict on Tencent HY3
HY3 is a strong free, agentic model — not a frontier coder or creative model, and it does not pretend to be. For building my Agent OS day to day I would still reach for GLM 5.2, Claude, Fable 5 or Opus 4.8.
As a free, spin-up-in-minutes option for agent workflows, it earns a place in the toolkit even if it is not a frontier model.
If GLM 5.2 and Qwen are my first picks among Chinese open-source models, HY3 earns its place as a fast, free option for agent tasks. I will run it across all 45 GoldieBench tasks next — and since it is free, you might as well test it yourself.
Inside, HY3 is already wired into Hermes and the HY3 Coder workflow, alongside GLM 5.2 and other open-source models — so you can plug it in and start building without the setup headache. You might also like my RTK vs Caveman comparison.
Yes — HY3 is free on OpenRouter until 21 July 2026, and also free in Hermes Agent (News Research) and Kilo Code.
What is Tencent HY3?
It is an open-source mixture-of-experts model from Tencent with ~295B total parameters (~21B active), designed for agent capabilities like browsing and tool use.
Is HY3 any good?
For free agentic tasks, yes — browsing, running tools and Remotion video work well. For frontier creative or game generation it is basic and glitchy compared with GLM 5.2, Fable 5, Claude or Opus 4.8.
Can you run Tencent HY3 locally?
Not easily yet — it was not on LM Studio at the time of testing. Local-running details are still emerging, though the active-parameter count keeps it efficient for its size.
HY3 vs GLM 5.2 — which is better?
For building and creative work, GLM 5.2 is nicer. HY3 is the better pick when you want a fast, free model purely for agent tasks.
The Bottom Line
Tencent HY3 is a genuinely useful free agentic model — grab it while it is free, lean on it for browsing, tools and Remotion, and keep a frontier model for the heavy creative lifting.
