Breaking

Tencent Releases 295B MoE Model Hy3 With 21B Active Params as Open Weights

July 7, 2026 at 11:15 EDT

Foundation Models
Open Source
AI Agents

Tencent's large language model team Hunyuan has released Hy3, a Mixture-of-Experts model with 295B total and 21B active parameters, as open weights under the Apache 2.0 license. Positioned as an efficiency-focused rival to trillion-parameter flagships, it targets agent-style tasks such as coding and tool use.

July 2026 · Tencent Hunyuan

Hy3: a 295B open model that fires only 21B per token

Tencent open-sources its Mixture-of-Experts model under Apache 2.0 — routing top-8 of 192 experts to rival far larger flagships while keeping per-token compute lean.

295B

Total parameters (MoE)

21B

Active per token

256K

Native context length

78%

SWE-bench Verified

Total capacity vs. what actually runs

The MoE design holds 295B parameters but activates only 21B per token — roughly a 14× gap between capacity and compute.

295B

Total capacity

21B

Active / token

Preview → Final: reliability gains

Large-scale post-training cut error rates and sharply raised long-context retrieval.

12.5%

5.4%

Hallucination

25.4%

12.7%

Commonsense err.

17.4%

7.9%

Multi-turn issues

42.9%

75.1%

MRCR retrieval ↑

Preview (error) Final (error, lower=better) Retrieval (higher=better)

In Tencent's own products

Integrating Hy3 into WorkBuddy, Yuanbao, WeChat and game assistants lifted task success rates.

72% → 90% task success rate

What developers like

Size efficiency: flagship-level from 21B active
Strong agentic & coding behavior
Apache 2.0 — commercial-friendly weights
Easy for debugging, reasoning and UI rework

Open questions

Preview showed mixed

Continue reading

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