Tencent releases Hy3, an open 295B-parameter MoE model with a 256K context window

Tencent's Hy team has released Hy3, an open Mixture-of-Experts (MoE) model with 295 billion parameters that activates only 21 billion parameters for each token it processes. The model is distributed under the Apache 2.0 license and is built for reasoning, agentic workflows and long-context tasks.

What is the architecture behind Hy3

Hy3 follows the Mixture-of-Experts approach: instead of using all 295 billion parameters at once, the system selects and activates only a portion for each token, about 21 billion parameters. This lets the model hold a large total body of knowledge while noticeably reducing compute load compared with 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 score: 78.0
• Free access on OpenRouter: until July 21, 2026

What the model is designed for

The developers position Hy3 as a model for reasoning, agentic workflows and long-context work, meaning tasks where the system has to keep a large amount of information in memory and carry out multi-step chains of actions. The 256K-token context window lets users feed the model large documents, codebases or long conversations without losing coherence.

On SWE-Bench Verified, a benchmark that measures how well models solve real code-fixing tasks in open repositories, Hy3 scored 78.0. The developers also claim a lower hallucination rate than the team's previous models.

How to try the model for free

Until July 21, 2026, Hy3 is available for free testing on OpenRouter, an API aggregator for large language models that lets developers call Hy3 without having to deploy the model locally.

What this means

Hy3 adds another large MoE model under the Apache 2.0 license to the open market, one that competes with leading proprietary systems on context length and coding benchmark results. It also lowers the barrier to entry for developers who need access to a powerful model without relying on closed APIs.

Frequently asked questions

What is Mixture-of-Experts in Hy3?

It is an architecture in which the model has 295 billion parameters in total but activates only about 21 billion for each token. The remaining "experts" are not used at that moment, which reduces compute costs.

How much does access to Hy3 cost?

Until July 21, 2026, the model is free on OpenRouter. The Apache 2.0 license also allows you to deploy the model yourself without license fees.

How does Hy3 perform on coding benchmarks?

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