Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

PulseOpen-Weights Releases

Tencent ships Hy4 flash open weights: 220B MoE, 12B active

X is buzzing over Tencent Hy4 flash's open-weights drop: a 220B MoE with 12B active parameters and 1M context, under a restrictive new license.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A single open workstation tower on a wooden desk in a dim home office, RAM sticks and a graphics card visible inside, an amber status light glowing on the motherboard — the kind of machine that serves a 12B-active open-weights MoE locally. Illustration: DeAI
A single open workstation tower on a wooden desk in a dim home office, RAM sticks and a graphics card visible inside, an amber status light glowing on the motherboard — the kind of machine that serves a 12B-active open-weights MoE locally. Illustration: DeAI

Tencent opened the weights of Hy4 flash, its efficient multimodal mixture-of-experts model, on Hugging Face, and X is treating it as the open-weights event of the day. The verified facts underneath the noise: a 220B-parameter model with 12B active per token, a 1M-token context window, and a license that is not Apache 2.0 — DeAI read the LICENSE file so builders don't have to guess.

Key facts

  • The model card specifies 220B total parameters (including 64B Engram parameters) with 11.3B text parameters active per token, and a context window of 1M (1,048,576) tokens.
  • Architecture per the card: a 42-layer language backbone (one dense layer plus 41 MoE layers), 16 of 512 routed experts activated per token, hybrid sparse and sliding-window attention, and a built-in 0.85B speculative-decoding drafter (DFly2).
  • DeAI read the LICENSE file: Hy4 flash ships under the Tencent Hy License v1.0, derived from Apache 2.0 with added MaaS commercial terms — a separate written license is required above US$50M gross revenue in any 12-month period or 100M monthly active users.
  • The weights landed November 17–18 per the repo timeline, two weeks after the Hy4 API launch (Reuters-covered, November 4); community GGUF quants followed within hours via AngelSlim/Hy4-flash-GGUF.
  • Velocity: the first widely shared post opened the day's conversation, with follow-on threads and Chinese-lab trackers spreading the specs within hours.

What's driving the conversation

The thread started with the weights-drop announcement and spec recaps, then forked into two camps. One camp treats Hy4 flash as the confirmation of the now-standard Chinese-lab playbook — API first, open weights after a safety-review window — the same sequence as GLM-5.3 and Kimi K3.1 Pro, with full Hy4 weights teased as next. The other camp is arguing about what "open" means here: several early posts described the release as Apache-style, which the LICENSE file does not support.

The strongest engagement clusters around the self-hosting math. A 12B-active model at 1M context is a different machine profile from a dense model of the same total size, and the discussion is mostly about what hardware actually serves it well and whether the community quantizations hold up. One widely shared thread recounts the spec sheet; the replies are largely hardware-strategy talk rather than benchmark discourse.

The substance

What is verified: the weights exist, the repo is live, and the architecture table on the model card is explicit. DeAI's pre-write check confirmed the Hugging Face repos for Hy4-flash and the FP8 variant, with the weights' last-modified timestamps consistent with a November 17–18 drop, and independent secondary coverage has since appeared in unite.ai, mlq.ai, and TechNode.

What is a claim, not yet a fact: every benchmark number, cost-efficiency ratio, and the "Engram parameter" framing in Tencent's model card and press materials. The model card's own benchmark section is the publisher's evaluation; no independent reproduction exists yet. Treat performance figures as vendor-attributed until someone else runs them.

What needed correcting before publication: the license. The X discourse repeatedly conflated it with the preview's Apache 2.0. DeAI verified the LICENSE file directly: Hy4 flash is released under the Tencent Hy License v1.0, which is derived from Apache 2.0 but adds Section 10 commercial terms — separate written licensing required for MaaS operations above US$50 million aggregate gross revenue in any consecutive 12-month period, or for any product with more than 100 million monthly active users. Internal use is explicitly exempt. That is a meaningful restriction for anyone considering serving it, and a change from the Hy4 preview's Apache 2.0 — worth stating precisely rather than glossing.

Why builders are watching

Three practical hooks. First, self-hosting economics: a 12B-active MoE changes the memory-versus-throughput trade-off for local serving, and the community GGUF quants already exist for llama.cpp-class runtimes. Second, license diligence: the Tencent Hy License's revenue carve-outs make it a different deployment decision from the Apache-2.0 open-weights releases DeAI has covered, like Kimi K3.1 Pro and MiniMax M3.1 Flash's community license — check Section 10 against your own revenue before building on it. Third, the pattern: if full Hy4 weights follow the same API-first-then-weights sequence, the open-weights frontier keeps widening on a predictable cadence, which is exactly what our coverage of Meituan's LongCat release argued matters more than any single checkpoint.

Questions

What did Tencent release in Hy4 flash?
Open weights for Hy4 flash, a natively multimodal mixture-of-experts model. Per the Hugging Face model card it has 220B total parameters including 64B Engram parameters, with 11.3B text parameters active per token, a 42-layer backbone (one dense layer plus 41 MoE layers, 16 of 512 routed experts activated per token), and a 1,048,576-token context window.
What license does Hy4 flash use?
The Tencent Hy License, Version 1.0 — verified by reading the repo's LICENSE file. It is derived from Apache 2.0 but adds commercial terms: a separate written license is required for model-as-a-service businesses above US$50 million in gross revenue over any 12-month period or products with more than 100 million monthly active users. Internal use is exempt.
What's disputed about the Hy4 flash release?
Benchmark scores, cost-efficiency ratios, and the 'Engram parameter' framing come from Tencent's own model card and press release and have no independent reproduction yet. Early X posts also described the license loosely; the actual terms are the $50M-revenue / 100M-MAU MaaS carve-outs above, not a blanket Apache 2.0.

Sources

  1. X post on the Hy4 flash weights drop (discourse driver) — X
  2. X post: ai_algoit on Hy4 flash specs — X
  3. X post: ChinaTechFlux on Hy4 flash — X
  4. tencent/Hy4-flash model card — Hugging Face
  5. tencent/Hy4-flash LICENSE file (Tencent Hy License v1.0) — Hugging Face

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →