Breaking

Tencent Ships Hy3, a 295B MoE Open Model, Under Apache 2.0

July 6, 2026 at 08:24 EDT

Foundation Models
Open Source
Infra & Chips

Tencent's Hy team on July 6, 2026 released Hy3, a 295B total-parameter open-weight model built on a Mixture-of-Experts (MoE) design. It ships under the Apache 2.0 license, with weights available across multiple platforms including Hugging Face, GitHub and ModelScope.

July 6, 2026 · Tencent

Hy3: an open 295B MoE model built for real work

Tencent's fully open Mixture-of-Experts model activates just 21B of its 295B parameters, ships under Apache 2.0, and is tuned to rival flagships 2–5× its size on agentic and productivity tasks.

295B

Total parameters
(21B active at inference)

256K

Context length,
tokens

50+

Internal use cases
feeding post-training

Apache 2.0

Open weights on HF,
ModelScope, GitCode

WorkBuddy: task resolution rate

Hy3 vs GLM-5.2 in internal evaluations — each block ≈ 10%.

72%

GLM-5.2

90%

Hy3

−34%

avg. completion time

−47.4%

doc-processing tokens

−49.0%

PPT-creation tokens

Reliability improved (before → after)

Hallucination rate12.5% → 5.4%

→

Common-sense error rate25.4% → 12.7%

→

Multi-turn issue rate17.4% → 7.9%

→

MRCR (higher is better)42.9% → 75.1%

→

Blind practical test · 270 experts

2.67/4

Hy3

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.