NewsBite analysis · AI, Cloud, Programming, Tools
MarkTechPost
An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context
Tencent has introduced Hy3, a Mixture-of-Experts model boasting 295 billion parameters but activating only 21 billion for efficiency. Hy3 aims to enhance reasoning and long-context workflows, delivering substantial improvements in core tasks. Additionally, its anti-hallucination training significantly reduces errors, marking a leap forward in AI reliability. A free trial of the model is available for an initial two weeks.
ExperimentConfidence: HighImmediate horizon
Published 7 Jul 2026, 06:40 UTC4 min read
Visit original article ↗Lead image — MarkTechPost
Summary
TL;DR
- Hy3 model activates only 21B parameters per token, optimizing performance.
- Designed for agentic workflows and long-context tasks.
- Hallucination rates reduced from 12.5% to 5.4%.
- Supports various coding and reasoning benchmarks with high accuracy.
- Free access available until July 21, 2026.
Interesting facts
- Hy3 encompasses a sparse architecture with 192 experts.
- It allows the activation of only 8 experts per token during computation.
- The model features enhancements for tool calling and output formatting.