Tencent Open-Sourced LLM Hy3 with 295 Billion Parameters
Tencent published the open-source code for LLM Hy3 — a Mixture-of-Experts (MoE) language model with 295 billion parameters, of which 21 billion are active during inference. The model has a 256K token context window. According to Tencent's claims, Hy3 surpasses open models of similar size and competes with flagship open LLMs that are 2–5 times larger. Apache 2.0 license allows free use of the model in commercial projects. Until July 21, Hy3 is available for free on the OpenRouter platform.
AI-processed from Simon Willison; edited by Hamidun News
Tencent released LLM Hy3 (Hyperbolic Model 3) under a free Apache 2.0 license. It is one of the few open models from the Chinese AI leader, provided to the community for development and research.
Hy3 Characteristics
The model is built on a Mixture-of-Experts (MoE) architecture and comprises 295 billion parameters. However, during inference, only 21 billion parameters are active, which reduces computational requirements without sacrificing performance. The MTP layer adds 3.8 billion parameters for output tuning. The context window length is 256K tokens — sufficient for analyzing long documents, codebases, or dialogs.
The full model size on disk is 598 gigabytes. The quantized version in FP8 format occupies 300 gigabytes, making it more accessible for deployment on standard hardware.
Performance
According to Tencent, Hy3 outperforms open models of similar parameter size in tests and shows results comparable to flagship models that are 2–5 times larger. The company gathered feedback from 50+ products and application makers, refining the model after the Preview version released in April 2026.
- Total parameters: 295 billion
- Active parameters during inference: 21 billion
- Context window: 256K tokens
- Full model size: 598 GB
- Quantized size (FP8): 300 GB
- License: Apache 2.0
Availability and Deployment
The model is hosted on Hugging Face and available for free use on the OpenRouter platform until July 21, 2026. This allows developers to test the model before deploying it to their own infrastructures. Apache 2.0 means that companies can use Hy3 in commercial products without restrictions, provided they comply with license conditions (mainly — attribution).
Why this matters
The release of Hy3 strengthens Tencent's position as a significant player in the global LLM market. Until now, most open and competitive models came from Western labs (Meta, Mistral, OpenAI) or other regions (Alibaba, ByteDance in China). Hy3, with its 256K context and MoE architecture, demonstrates that Chinese developers have achieved parity in scale and model capabilities.
For the developer community, this expands the choice of open models and allows more flexible tool selection for specific tasks without dependence on closed APIs.
What this means
The global market for open LLMs is becoming more competitive, which reduces API and infrastructure costs and accelerates innovation. Companies like Tencent invest in open models as a strategy to gain community trust and strengthen their brand in AI. Free access on OpenRouter until mid-July is a marketing move that allows rapid user acquisition and feedback collection.
Want to stop reading about AI and start using it?
AI News is a curated feed of AI/tech news. Hamidun Academy teaches you to use AI systematically in your work.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.