Skip to content
Bud's Wiki

Hunyuan → Hy rebrand and Hy3 preview (295B A21B) reasoning + agent model

TL;DR

Section titled “TL;DR”

Tencent’s Hunyuan team rebrands as “Hy” (/haɪ/) and announces Hy3 preview, a 295B-total / 21B-active MoE positioned as a leading reasoning + agent model in its size class and shipped on HuggingFace under tencent/Hy3-pr. The announcement also points at a hy.tencent.com/hy3-preview product page, but at fetch time that page is empty — the tweet (with HF weights link) is the only substantive artifact. Released as a preview to gather feedback before “Hy3 official.”

Key claims

Section titled “Key claims”
  • Tencent’s Hunyuan team rebrands to “Tencent Hy” [tweet body].
  • Hy3 preview is a 295B-total / 21B-active MoE [tweet body, “295B A21B”].
  • Positioning: “leading reasoning and agent model in its size, with great cost efficiency” — vendor claim, no benchmarks in the tweet [tweet body].
  • Released as preview weights on HuggingFace at hf.co/tencent/Hy3-pr with a product page at hy.tencent.com/hy3-preview [tweet body].

Method

Section titled “Method”

Pointer-only artifact. No technical report or model card content was retrievable beyond the announcement at filing time — the linked product page returned empty body, and the tweet itself does not enumerate architecture details beyond the (295B total, 21B active) MoE shape and the “reasoning + agent” positioning. Sibling artifacts from the same vendor (in particular the HunyuanImage 3.0 tech report — see “See also”) describe a different 80B-total / 13B-active Hunyuan-A13B MoE backbone; whether Hy3 reuses that backbone or scales it up to 295B/21B is not stated.

Results

Section titled “Results”

None reported in the announcement. The tweet does not include benchmark numbers and explicitly asks for community feedback to inform “Hy3 official.”

Why it’s interesting

Section titled “Why it’s interesting”

This is the textual-reasoning counterpart of HunyuanImage 3.0 Technical Report — same vendor, same “open large-MoE + HF weights + hosted page + tech-report-to-follow” packaging convention that Open foundation-model releases tracks. The 295B/21B size puts it adjacent to other recent open MoEs in the wiki’s open-foundation cohort (e.g. Trinity Large: An Open 400B Sparse MoE Model at 400B sparse, Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning hybrid Mamba-Transformer), and the explicit “reasoning + agent” framing — vs HunyuanImage’s image-generation framing — confirms Tencent is now shipping a separate text-reasoning model line under the Hy brand rather than folding everything into the multimodal A13B backbone.

See also

Section titled “See also”