Breaking

Tencent Releases 295B MoE Model Hy3, Free on OpenRouter Until July 21

July 7, 2026 at 00:14 EDT

Foundation Models
Open Source
AI Agents

Tencent Hunyuan open-sourced Hy3, a 295B-parameter Mixture-of-Experts (MoE) model, on July 6, 2026 under the Apache 2.0 license. Built with a focus on coding, reasoning and agentic use, a free tier (tencent/hy3:free) is available on OpenRouter through July 21.

Early July 2026 · Tencent Hunyuan

Hy3 goes open source: a 295B MoE that punches at trillion scale

Tencent's final Hunyuan 3 build ships under Apache 2.0 — tuned for coding, reasoning and agents, with a limited free window on OpenRouter through July 21.

295B

Total parameters (MoE)

~21B

Active params at inference

256K

Native context tokens

90%+

Agent task resolution (claimed)

Sparse by design: total vs. active parameters

Top-8 of 192 experts fire per token — so only a fraction of the model is active at inference.

295B

Total

~21B

Active

~14× fewer active parameters — rivals flagships two-to-five times its size.

At a glance

License

Apache 2.0 (commercial-friendly)

Weights

BF16 ~598GB / FP8 ~300GB

Runtimes

Tencent Cloud, vLLM, llama.cpp

Free window

OpenRouter through 2026/7/21

The preview already led OpenRouter

Before the final build, Hy3's preview topped weekly token throughput — a signal of fast agent-workload adoption.

3.66T

Weekly tokens served

15.4%

Market share (#1)

50+

Product integrations fed feedback

What's landing well

Smooth real-world code fixing in hands-on reports
On par with DeepSeek V4 Pro on agentic benchmarks
Runs on a single DGX Spark (Q2_K) at 18.1 tok/s
Strong cost efficiency for its size

To confirm

Detailed benchmarks & pricing on the official page
llama.cpp support only recently added — bring-up bugs
Best-in-class claims still vendor-stated

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.