Tencent released Hy3—an open MoE model with 295 billion parameters and 256K context

The Tencent Hy team released an open model Hy3 — a Mixture-of-Experts (MoE) system with 295 billion parameters that activates only 21 billion parameters for each token processed. The model is distributed under the Apache 2.0 license and is designed for reasoning tasks, agentic scenarios, and long-context applications.

What is Hy3's architecture?

Hy3 is built on the Mixture-of-Experts principle: instead of engaging all 295 billion parameters of the model at once, the system selects and activates only a portion for each token — approximately 21 billion parameters. This approach allows the model to maintain a large total volume of knowledge while significantly reducing computational load compared to a dense model of the same scale.

• Total model size — 295 billion parameters (MoE)
• Active parameters per token — 21 billion
• Context window — 256,000 tokens
• License — Apache 2.0 (open source)
• SWE-Bench Verified result — 78.0
• Free access on OpenRouter — until July 21, 2026

What tasks is the model designed for?

The developers position Hy3 as a model for reasoning, agentic scenarios, and long-context work — that is, for tasks where the system needs to retain a large amount of information in memory and execute multi-step chains of actions. The 256K token context window allows the model to process large documents, codebases, or long dialogues without loss of coherence.

On the SWE-Bench Verified benchmark, which evaluates models' ability to solve real code-fixing tasks in open repositories, Hy3 achieved a score of 78.0. The developers also claim reduced hallucination frequency compared to previous models from their team.

How to try the model for free

Until July 21, 2026, Hy3 is available for free testing on the OpenRouter platform — an API aggregator for large language models through which developers can access Hy3 without needing to deploy the model locally.

What this means

The release of Hy3 adds another major MoE model with an Apache 2.0 license to the open market, competing in context and code benchmark results with leading proprietary systems, and lowers the entry barrier for developers who need access to a powerful model without closed APIs.

Frequently asked questions

What is Mixture-of-Experts in Hy3?

This is an architecture where the model has 295 billion parameters in total, but activates only about 21 billion for each token — the remaining "experts" are not engaged at that moment, reducing computational costs.

How much does Hy3 access cost?

Until July 21, 2026, the model is free on the OpenRouter platform; the Apache 2.0 license also allows for self-deployment without license fees.

What is Hy3's performance on code benchmarks?

On SWE-Bench Verified, a test for fixing real bugs in open repositories, the model scored 78.0 points.