Tencent has officially released Hy3, a Mixture-of-Experts model, and made it free on the LLM routing service OpenRouter through July 21. With 295B total parameters and a context window of up to 256K, it is built for coding, reasoning and agent use.
Open-Weight AI · Tencent Hunyuan
Hy3: a 295B open model that punches far above its active size
Tencent's new Mixture-of-Experts model wakes just 21B of its 295B parameters per token, ships free on OpenRouter through July 21, and claims agent performance rivaling far larger systems — all under a commercial Apache 2.0 license.
295B
Total parameters (MoE, 192 experts / top-8)
256K+
Native context window (up to 262K tokens)
73.2%
GPQA Diamond score claimed
Only 21B of 295B parameters run per token
The sparse MoE design keeps ~93% of the network dormant at inference — a fraction of the compute for the capability of a huge model.
Hallucination rate cut by more than half
Reported error rate on the final release versus earlier behavior — a key driver of its reliability on long agentic workflows.
Where it shines
Smooth, effective coding and bug fixing
Stable tool calling and long-running agent tasks
Claude Opus 4.8-class agent performance at lower cost
Runs locally at ~18.1 tok/s on a DGX Spark (Q2_K)
Watch-outs
Local deploy leans heavily on quantization + hardware
Multiple GPUs recommended for the large MoE
Much long-term data still comes from the preview
Adoption hinges on paid pricing after the free period
Positioned among China's open frontier
A commercially friendly, agent-focused open model lined up against DeepSeek V4 Pro and GLM-5.2 — and folded into Tencent's own Yuanbao and WorkBuddy.
Apache 2.0 · commercial use
3 reasoning modes: no-think / low / high
~$0.063 / M input tokens
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…