Tencent Hy3 is FREE right now (agentic model, 295B MoE) — here's how to use it before July 21
I've been testing Tencent's new Hy3 model for the past few days after plugging it into Hermes Agent for video editing work, and honestly it surprised me. Sharing the details here since I've seen a bunch of people asking "how to use Hy3 for free" and the info is scattered across random blogs.
What is Tencent Hy3?
Hy3 is Tencent's new open-source model, released July 6, 2026:
- 295B total parameters, Mixture-of-Experts (21B active, 192 experts, top-8 routing)
- 256K context window
- Apache 2.0 license (fully open weights on Hugging Face)
- Built specifically for agentic workflows — tool calling, browsing, multi-step tasks, long-context document work
- Configurable reasoning modes: no-think (fast) / low / high (for math, coding, complex tasks)
Where you can use it for free right now
- OpenRouter — search
Hy3or use the routetencent/hy3:free. Free until July 21, 2026 (this is a promotional window, not permanent — grab an API key and use it before then). - Hermes Agent (Nous Research) — free access built in.
- Kilo Code — also offering it free.
How it actually performs (Terminal-Bench 2.1)
This is the benchmark that matters most for agentic/terminal work, and it's where Hy3 genuinely stands out among open models:
| Model | Terminal-Bench 2.1 (pass@1) |
|---|---|
| Hy3 (Tencent, open source) | 71.7% |
| Qwen 3.7 Max (open weights) | 71.5% |
| DeepSeek V4 Pro (open weights) | 64.0% |
| GLM 5.1 (open weights) | 63.5% |
Hy3 basically ties Qwen 3.7 Max at the top and clears DeepSeek V4 Pro and GLM 5.1 by a solid 7-8 points — which is notable since Hy3 only has 21B active params, way less than what DeepSeek V4 Pro and GLM 5.1 are running. That efficiency-to-performance ratio is really the story here, not just "it's free."
Important: what happens after July 21
A lot of posts are vague about this, so to be clear — after the free window ends, Hy3 switches to paid, pay-per-token pricing, not unlimited-for-$5. It's currently listed at $0.14/M input, $0.58/M output. Worth budgeting for if you plan to keep using it past the promo window.
Where it's strong vs where it's not
Strong: agentic tasks, tool use, browsing, long-context document/financial analysis, coding agents, stable multi-step workflows. It even beats GLM 5.2 on the Claw Eval agentic benchmark.
Weak: frontier creative work. Multiple hands-on reviewers found outputs for game generation and complex frontend/creative prompts came out basic or glitchy compared to GLM 5.2, Claude, or Fable 5. If you need pure agent muscle on a budget, it's great. If you need frontier-level creative output, it's not there yet.