Tencent drops Hy3 under Apache 2.0
Tencent has released Hy3, an open-weights 295B Mixture-of-Experts foundation model under the Apache 2.0 license, optimized for reasoning, coding, and agentic workflows. The model features 21B active parameters, a 256K context length, and supports speculative decoding via a built-in 3.8B Multi-Token Prediction layer.
Tencent’s Hy3 enters the competitive open-weights arena as a direct challenger to GLM and DeepSeek, leveraging a massive Mixture-of-Experts (MoE) architecture with speculative decoding to provide high-quality reasoning at a lower active parameter cost.
* The 295B total parameters with only 21B active parameters represents a highly optimized routing ratio (192 experts, top-8 routing), balancing capacity and execution speed.
* A native 3.8B Multi-Token Prediction (MTP) layer enables faster speculative decoding out of the box, reducing inference latency for long-context and code-generation tasks.
* Apache 2.0 licensing makes this model highly appealing for enterprise deployments and custom finetuning without proprietary restrictions.
* Heavy hardware requirements (e.g., 8x H20 or H200 GPUs) mean deployment is mostly geared towards enterprise settings rather than consumer hardware.
DISCOVERED
1h ago
2026-07-06
PUBLISHED
1h ago
2026-07-06
RELEVANCE
AUTHOR
Bijan Bowen
