The Field Guide to Fable
Tencent released Hy3, a 295B parameter Mixture-of-Experts (MoE) model with 21B active parameters, licensed under Apache 2.0. The architecture features 192 experts with top-8 routing, Grouped Query Attention (GQA), 256K context window, and a 3.8B MTP layer for speculative decoding. Hy3 demonstrates day-0 maturity with native support in vLLM, including upstreamed production kernels for load-balanced decode scheduling and fused FP8 MoE serving. Performance optimizations include up to 2.95x speedup
75
Hot
65
Quality
70
Impact
Analysis
Disclaimer: The above content is generated by AI and is for reference only.
LLM Research Product Launch
Related Articles
GPT-5 Pro Self-Proves Mathematical Theorem: Has AI's PhD-Level Moment Arrived?
The Person That $2.7 Billion Couldn't Retain, Altman Waited Ten Years
[GitHub] apache/texera
OpenAI’s Turbulent Week: GPT-5.6 Launch, Executive Departure, Legal Setback and Product Shifts Mark Pivotal Stretch for the Company
Bhavin Turakhia Bootstraps $30M Enterprise AI Platform Neo, Betting on Ground-Up Redesign of Workplace Software