Tencent Releases Hy3: 295B MoE Model with 256K Context and Configurable Reasoning Modes
Tencent has released Hy3, a 295-billion parameter Mixture-of-Experts model with 21 billion active parameters and a 256,000-token context window. The model features configurable reasoning modes and is available free through OpenRouter, with deployment ending July 21, 2026.
Tencent Releases Hy3: 295B MoE Model with 256K Context and Configurable Reasoning Modes
Tencent has released Hy3, a 295-billion parameter Mixture-of-Experts (MoE) model with 21 billion active parameters and a 256,000-token context window. The model is available free through OpenRouter until July 21, 2026.
Architecture and Capabilities
Hy3 uses a sparse MoE architecture with 192 experts and top-8 routing, activating 21 billion parameters per forward pass. According to Tencent, the model supports three reasoning modes: a default direct "no-think" mode for standard queries, plus low and high chain-of-thought modes for complex tasks involving mathematics, coding, and multi-step reasoning.
The 256K context window targets long-horizon tasks including document processing, multi-turn conversations with constraint tracking, and extended agentic workflows. Tencent claims the model provides improved coreference resolution and stable tool-calling that generalizes across different agent frameworks.
Design Philosophy
Tencent positions Hy3 as focused on "grounded, anti-hallucination behavior," claiming the model is designed to flag when evidence is missing rather than fabricate responses. The company states this makes it suitable for production use cases including coding, financial analysis, game development, and frontend design.
No benchmark scores, training data details, or independent performance metrics have been disclosed. The model architecture uses 192 experts with top-8 routing, meaning each token is processed by 8 of the 192 available expert networks.
Availability
Hy3 is available free through OpenRouter's API, which provides OpenAI-compatible endpoints. The free tier runs through July 21, 2026, after which pricing or availability terms have not been announced.
The model's release date is listed as July 6, 2026, though this appears to be listed incorrectly as a future date. OpenRouter provides monitoring for uptime, throughput (tokens per second), latency, and time-to-first-token metrics, though specific performance numbers were not included in the release.
What This Means
Hy3 represents Tencent's entry into the large-scale MoE model space, competing with models like Mixtral and DeepSeek-V3. The configurable reasoning modes are notable, allowing developers to trade off speed versus reasoning depth. However, without published benchmarks or independent evaluation, it's difficult to assess how Hy3 compares to existing models. The limited free availability window suggests this may be a pilot or promotional release rather than a long-term deployment strategy. The 256K context window puts it in competition with models like Claude 3.5 Sonnet and GPT-4 Turbo for long-document processing tasks.
Related Articles
Tencent Releases Hy3: 295B-Parameter MoE Model with 21B Active Parameters at 256K Context
Tencent has released Hy3, a 295-billion parameter Mixture-of-Experts model with 21 billion active parameters and 3.8 billion MTP layer parameters. The model features a 256K context window and is released under Apache 2.0 license, with pricing not yet disclosed.
DeepSeek Releases V4 Models: 1M Context Window, 90% Less KV Cache Than V3
DeepSeek has released two new MoE models: DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated). Both models support a one million token context window and use a hybrid attention architecture that requires only 27% of single-token inference FLOPs and 10% of KV cache compared to DeepSeek-V3.2.
DeepSeek Releases V4-Pro with 1.6T Parameters, 1M Token Context at 27% Inference Cost of V3
DeepSeek has released two Mixture-of-Experts models: V4-Pro with 1.6 trillion parameters (49B activated) and V4-Flash with 284B parameters (13B activated), both supporting 1 million token context windows. V4-Pro requires only 27% of inference FLOPs and 10% of KV cache compared to V3.2 at 1M token context, trained on over 32 trillion tokens.
Mistral releases Leanstral 1.5: 119B parameter open-source model for Lean 4 proof assistance
Mistral AI has released Leanstral 1.5, an open-source 119B parameter mixture-of-experts model designed specifically for Lean 4 proof assistance. The model features 128 experts with 4 active per token (6.5B activated parameters), a 256k token context window, and multimodal input capabilities.
Comments
Loading...