Tencent open-sourced Hy3 this week. OpenRouter is running it for free until July 21. If you want to test a competitive new model without spending anything, this is the window.

Free link: https://openrouter.ai/tencent/hy3:free

What is Hy3

A 295B parameter Mixture-of-Experts model. Only 21B parameters activate per request. 256K context window.

In plain terms: a large model that runs cheap because it only uses a small slice of itself for each task.

Why try it

I tested it on coding and multi-step tasks. It holds up. On SWE-bench Pro, BrowseComp, and MCP Atlas, it trades scores with GPT 5.5, Claude Opus 4.8, and DeepSeek V4 Pro, all models with 2-5x its active parameter count.

Tencent also claims the hallucination rate dropped from 12.5% to 5.4% between Preview and this release. Multi-turn error rate went from 17.4% to 7.9%. I can't verify those internal numbers, but in my testing, the outputs felt grounded. Tencent built it with anti-hallucination as a priority: it answers when grounded, flags when evidence is missing.

How to use it

On OpenRouter, select tencent/hy3:free. No API key billing, no credit card.

In code:

model: tencent/hy3:free

The model supports three reasoning modes:

  • no_think (default): direct response, fast
  • low: light chain-of-thought
  • high: full reasoning, use this for math or complex coding

Caveats

  • Free tier has rate limits. OpenRouter doesn't guarantee uptime or capacity for free models.
  • The free window ends July 21. After that, standard pricing applies.
  • 295B total params means self-hosting needs serious hardware (8x H20 or similar). Most people will use it through an API.
  • It's new. Community feedback is still thin. Run your own tests before putting it in a production workflow.