Tencent Releases Hunyuan Hy3, an Apache 2.0 Open-Weight Model Tuned for Agentic Tasks
The 295B MoE topped Hugging Face trending in its launch week and was free on OpenRouter through July 21 — though hands-on reviews find its coding below the frontier and its headline claims vendor-reported.
Tencent released Hunyuan Hy3 on July 6, 2026, an open-weight model tuned for agentic workloads under an Apache 2.0 license, per the hokai.io model page.
Context
Hy3 joined a dense early-July cluster of Chinese open-weight releases — Meituan's LongCat-2.0 (July 5) and Kimi K2.7 Code's arrival in GitHub Copilot (July 1) — ahead of Moonshot's Kimi K3 on July 16. Tencent pitched agentic task performance rather than raw frontier capability.
What changed
- Specs: 295B-total MoE with roughly 21B active parameters, 256K-token context, 32,768 max output, text and image input, Apache 2.0.
- Distribution: #1 on Hugging Face trending the week of July 5–11; free via OpenRouter through July 21.
- Positioning: agentic-tool use and multi-step task execution over benchmark-chasing.
Why it matters
Tencent — one of China's largest platform companies — committing its agent model to Apache 2.0 deepens the permissively licensed Chinese open-weight stack that Western inference providers and enterprises are increasingly routing workloads to. Free OpenRouter access through July 21 was an adoption accelerant for the exact developer audience GitHub Copilot courted with K2.7 Code days earlier.
Details
Hands-on reviews summarized in the research record find real agentic strength but coding performance below GLM-5.2 and Fable 5. Tencent's claims of halved hallucination rates and a 78.0 SWE-bench score are vendor-reported; one hands-on review found output quality below what the benchmark table implies.
Competitive frame
Hy3's 295B/~21B-active profile targets efficient serving rather than trillion-parameter scale, putting it in a different deployment class from LongCat-2.0 or Kimi K3 — runnable by well-resourced inference providers rather than only supernode operators. That efficiency focus, plus Apache 2.0 licensing, makes it a candidate default for agent pipelines that need tool-use competence without frontier pricing. Its trending-topping launch week suggests the distribution strategy worked, though trending ranks measure attention, not retained usage.
Limitations and caveats
Headline quality claims are vendor-reported and not independently verified; the research record flags the hallucination-halving claim specifically as awaiting verification. Free-access windows can inflate early adoption signals.
Sources
- hokai.io — Hunyuan Hy3 model page (aggregator/model hub)
*Update note: This post was last reviewed on 2026-07-22. Free OpenRouter access ended July 21; watch for post-promotion usage data.*
Sources
- hokai.io — Hy3 model page — aggregator
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.