Another giant model landed on Hugging Face, and this one is worth stopping for. On July 6 Tencent open-sourced Hy3, a 295B mixture-of-experts model under a plain Apache 2.0 license, and the one-line version we'd give a colleague is this: it's genuinely open, genuinely large, and only lights up 21B of its 295B parameters on any given token. It's free to try on OpenRouter through July 21. It also isn't running on your gaming PC, and that gap is the whole story. On paper it edges past GLM-5.1 in Tencent's own testing and posts strong reasoning scores. So the question isn't whether it looks good. It's whether it earns a slot in your stack.
The short answer
Tencent open-sourced Hy3, a 295B mixture-of-experts model with 21B active parameters and 256K context, under a permissive Apache 2.0 license. It posts strong reasoning scores and, by Tencent’s own blind test, edges past GLM-5.1. The catch is hardware: Tencent points at 8 big GPUs to serve it. Free to try on OpenRouter right now, free to self-host if you have the metal.
What Tencent actually shipped
Weights, and a license that matters. On July 6 the Hunyuan team put Hy3 on Hugging Face (also ModelScope, GitCode and CNB) under Apache 2.0. That last part is the news as much as the size. A 295B model you can download, fine-tune and ship commercially with no gate is not something you get every week, and it lifts the geographic limits that hemmed in the April preview.
The shape of it: 295B total parameters, 21B active. It’s a mixture-of-experts, so each layer holds 192 routed experts plus one that’s always on, and a router picks the top 8 per token. 80 layers, a 256K context window, BF16 weights, and an FP8 checkpoint for teams who want a smaller footprint. There’s also a 3.8B multi-token-prediction layer bolted on, which lets the model guess more than one token per step so decoding runs faster. Recommended sampling, if you’re wiring it up, is temperature 0.9 and top_p 1.0.
The benchmarks, and how far to trust them
Hy3’s own scorecard is genuinely strong on hard reasoning. GPQA Diamond at 90.4. USAMO 2026 at 72.0. IMOAnswerBench at 90.0. HLE, with tools, at 53.2. Those are the kind of numbers that used to belong to closed flagships, and they’re coming from open weights you can pull tonight.
The comparison people will care about is GLM-5.1, the other strong open Chinese model. Tencent ran a blind test: 270 experts, 312 real workflow comparisons, and Hy3 came out at 2.67 out of 4 against GLM-5.1’s 2.51, strongest in frontend and CI/CD work. It also claims a real reliability jump over its own April preview, with hallucinations down from 12.5% to 5.4% and commonsense errors from 25.4% to 12.7%.
Now the honest part. Every one of those figures is Tencent’s own, published alongside the model it’s selling. That’s normal, and it doesn’t make them wrong, but a vendor’s blind test is a starting point, not a verdict. We’d wait for independent runs (and honestly, run our own on the workloads we actually care about) before calling it a GLM-5.1 killer. If you’re weighing the open Chinese models against each other, our look at GLM-5.2 versus GPT-5.5 and Opus 4.8 lays out how these benchmark claims tend to hold up once you pressure them.
The catch: what it takes to run
Here’s where “open” stops meaning “yours”. Tencent’s own guidance is to serve Hy3 on 8 GPUs with large memory, naming the H20-3e class, and it ships the FP8 variant precisely because BF16 at this size is brutal on VRAM. It runs on vLLM (with the MTP layer doing speculative decoding) or SGLang (with EAGLE). Real inference stack, real data-center hardware. Not a homelab afternoon.
So for most of us, “using Hy3” this month means the free OpenRouter endpoint,
tencent/hy3:free, which runs through July 21 before standard pricing kicks
in. That’s the low-friction way to see whether its answers actually beat what
you’re already paying for. If your interest is running an open model on your
own box, Hy3 is the wrong size for that, and something like
Qwen 3.7 offline is a far more honest fit
for a single machine.
The honest read
Hy3 is a real gift to anyone building on open weights, and the Apache 2.0 license is the part that’ll still matter in a year. A 295B MoE with 21B active is a smart design, the reasoning scores are legitimately high, and the free window means you can judge it yourself before the pricing meter starts. What we wouldn’t do is treat the GLM-5.1 comparison as settled, or assume “open” means you’ll be running this on your own metal any time soon. Pull it up on OpenRouter this week, throw your own hard prompts at it, and see if it holds. That’s the only benchmark that pays your bills.
Sources: Tencent’s official Hy3 repository and model card on Hugging Face, with reporting and specs via MarkTechPost and Simon Willison, July 2026. Benchmark and blind-test figures are Tencent’s own and are not yet independently reproduced.
Frequently asked questions
What is Tencent Hy3?
Hy3 is Tencent Hunyuan's open large language model, released July 6, 2026. It is a 295B-parameter mixture-of-experts (MoE) model with 21B active parameters per token, a 256K context window, and Apache 2.0 licensed weights. Tencent aims it at reasoning and agent work, and ships both BF16 and FP8 checkpoints.
Is Hy3 really open source and free to use commercially?
The weights are released under Apache 2.0, which lets you download, self-host, fine-tune and ship the model in commercial products without paying a license fee. That is a genuinely permissive license, not a research-only or gated one. You still supply your own hardware to run it, or pay a host that serves it.
Can I run Hy3 locally?
Not on a normal desktop. Tencent recommends serving it on 8 GPUs with large memory, naming the H20-3e class, and ships an FP8 variant to cut the footprint. It runs on vLLM or SGLang. For a laptop or a single-GPU box, a much smaller open model is the realistic choice, not Hy3.
Is Hy3 better than GLM-5.1?
In Tencent's own blind evaluation, 270 experts scored Hy3 at 2.67 out of 4 against GLM-5.1 at 2.51 across 312 real workflow comparisons, with the biggest gains in frontend and CI/CD tasks. That is a self-reported test from the model's maker, so treat it as a promising signal and verify on your own workloads before trusting it.
How much does Hy3 cost?
The weights are free under Apache 2.0. To try it without hardware, OpenRouter runs a free endpoint at tencent/hy3:free through July 21, 2026, after which standard per-token pricing applies. Self-hosting has no license cost but does carry the GPU and power bill of serving a 295B model.