Tencent Releases Hy4 Flash Open Weights: 224B Parameters, 12B Active, and a 1-Million-Token Context
Tencent has published the open weights for Hy4 Flash, its cost-focused Hunyuan model, two days after the API launch. The 224-billion-parameter model activates only 12 billion parameters per token, handles text and images across a 1-million-token context, and is aimed squarely at everyday office and agent work.
Tencent Releases Hy4 Flash Open Weights: 224B Parameters, 12B Active, and a 1-Million-Token Context
Tencent has opened up Hy4 Flash, the efficiency-focused model in its Hunyuan family, publishing the weights for anyone to download, inspect and run on their own hardware. The release landed on Hugging Face on November 17, two days after the model first appeared as a paid API on November 16, and it arrives at a moment when the race in open-weight AI is being fought as much on running costs as on raw capability.
Hy4 Flash is not a shrunken flagship. Tencent describes it as a knowledge-distilled derivative of Hy4, rebuilt around one question: how cheaply can a model handle the high-frequency work people actually do every day — drafting documents, fixing spreadsheets, building slide decks, researching, and driving agents — without giving up too much quality. The answer, on Tencent's numbers, is a model with 224 billion total parameters that activates only about 12 billion of them for each token it generates.
Three versions, one download
The Hugging Face release comes in three checkpoints. The standard Hy4 Flash is the instruction-tuned chat model, complete with its reasoning modes. Hy4 Flash-FP8 is the same model quantised to 8-bit floating point for production serving, and Hy4 Flash-Base is the pre-trained checkpoint with no post-training at all, aimed at teams that want to fine-tune it themselves. The weights are also mirrored on GitHub, ModelScope and GitCode, with serving and fine-tuning code in the official GitHub repository.
The model accepts text and image input and supports a context window of up to 1,048,576 tokens — a million tokens, in round terms — which puts long documents, large codebases and extended agent sessions comfortably within reach.
Built for everyday work at a fraction of the cost
Tencent's pitch is cost per task rather than leaderboard position. On WorkIndex, the company's internal benchmark built from typical WorkBuddy office tasks such as slide creation, document editing, spreadsheet repair and audio and video processing, Hy4 Flash completed work at one-twentieth the per-task cost of the full Hy4 model, one-thirteenth the cost of Hy4 Preview, and one-eighth the cost of DeepSeek-V4.1-Flash, according to Tencent. It also slightly outscored the much larger Hy4 Preview — a 770-billion-parameter model with 49 billion active parameters — on overall effectiveness in the same internal evaluation.
A blind test with 85 internal WorkBuddy users across 259 tasks in 27 categories gave Hy4 Flash a satisfaction score of 2.79 out of 4, close to Claude Haiku 5.5's 2.81 and ahead of GLM-5.3-Flash at 2.53, Tencent said. These are the company's own figures, and no independent evaluation had been published at the time of writing, so they should be read as Tencent's account of its own model rather than a settled comparison.
On the hosted API, Hy4 Flash is priced at 0.070 US dollars per million input tokens, 0.281 dollars per million output tokens and 0.003 dollars per million cached input tokens. It is free on WorkBuddy until November 30, and available at half price through Tencent Cloud's TokenHub until then. The model is also live in Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter integration on the way.
The engineering behind the efficiency
Three architectural choices do most of the work. Hybrid attention mixes sliding-window attention with sparse attention to shrink the memory the model needs to hold while it generates text; Tencent says the key-value cache is one-sixth the size of Hy4's and one-twenty-fourth the size of Hy3's. LatentMoE, a mixture-of-experts design, is what keeps only around 12 billion parameters active per token. And Engram, an n-gram embedding system built from multi-head hash tables, stores roughly 63.5 billion parameters of learned word-sequence representations that are looked up rather than computed — the single largest contributor to the 224-billion total.
On the inference side, Tencent combines its megakernel optimisations, 4-bit quantisation and a five-layer draft model called DFly2 for speculative decoding, reaching decoding speeds above 150 tokens per second, according to the company. Post-training reinforcement learning cut average output tokens per task by half, Tencent said.
One small discrepancy is worth noting for anyone reading the model card: Tencent's launch material describes the model as 224 billion parameters, while the model card's own component table sums to about 220.3 billion — a 155.5-billion backbone, 63.5 billion Engram, plus small vision and draft-model components. The difference appears to come from how the draft model and rounding are counted, and it changes nothing about the model itself.
Read the licence before downloading
The weights are released under the Tencent Hy License, Version 1.0, an Apache 2.0 derivative with an added section that matters for commercial users. If you operate a model-as-a-service business with worldwide gross revenue above 50 million US dollars in any twelve-month period, or a product with more than 100 million monthly active users, you need a separate written licence from Tencent before using the weights. Internal use is exempt. It is a narrower grant than a plain Apache licence, and anyone planning a hosted service should read the licence file in the GitHub repository first.
Why it matters
Hy4 Flash continues a pattern that has defined the last few weeks in open-weight AI: flagship models get the headlines, then cheaper derivatives do the actual work. Alibaba released Qwen4-Max, its new flagship model, on November 16, the same day Hy4 Flash launched as an API, and Google has been pushing its own efficiency tier with Gemini 4 Hydrogen. Tencent's bet is that the market for a model that is good enough for everyday work at a twentieth of the cost is bigger than the market for the most capable model money can buy — and by opening the weights, it is inviting developers to test that bet on their own machines.
Tencent has also said the flagship Hy4's weights will follow. For now, Hy4 Flash is the company's open-weight offering, and its real test will be whether independent developers and evaluators find the same balance of cost and quality that Tencent's own benchmarks promise.
Sources
- HPCwire — Tencent Launches Hy4 Flash to Power High-Speed, Cost-Efficient AI Deployments (Tencent announcement, November 18, 2026)
- OrcaRouter — Hy4 Flash Open Weights: Read the Licence Before You Download
- AlphaSignal — Tencent Opens Hy4 Flash Weights With 1M-Token Context at 20x Lower Cost
- ToolNavs — Tencent's Hunyuan Hy4 Flash Goes Open-Weight: a 224B Lightweight Model, Free to Download
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Angry
0
Sad
0
Wow
0