Tencent Hy4-flash Card Says Weights Are Out; Launch Coverage Says Planned
A.I. / news
Tencent Hy4-flash Card Says Weights Are Out; Launch Coverage Says Planned
The Hugging Face page lists 220B parameters and a custom licence, while launch coverage gives 224B and a future release, at $0.070 per million input tokens.

The Hugging Face page for Tencent's Hy4-flash lists the model as a downloadable release with 220 billion parameters, while launch coverage dated November 16 says the weights are still planned and puts the size at 224 billion. As of Thursday evening, the two accounts had not been reconciled in anything The Terminal reviewed.
The model card on Hugging Face shows instruct, FP8 and base variants, no gating notice and a licence named the Tencent Hy License. Its news section says the models were released on Hugging Face, ModelScope and GitCode. It gives no release date.
220B or 224B, and what active means
The card lists 220B total parameters, of which 64B are Engram parameters, and 12B active per token. Excluding Engram, the vision encoder and the drafter, it lists 155.5B. The 12B active figure splits into 11.3B for text and 0.43B for vision.
The architecture is a mixture of experts with 42 layers, 512 routed experts, 16 active per token and one shared expert. Context is 1,048,576 tokens, with text and image input and text output. The card's vLLM example runs the FP8 variant with 8-way tensor parallelism and states no memory requirement.

| Item | Hugging Face card | Launch coverage |
|---|---|---|
| Total parameters | 220B | 224B |
| Active per token | 12B | 12B |
| Weights | Released | Planned |
| Context | 1,048,576 tokens | 1M tokens |
The AF.NET write-up gives the launch as November 16 and says open-weight release is planned, citing Unite.AI. It also notes that a headline on MLQ.ai says Tencent released open weights at 220B, which conflicts with Unite.AI. The nocode.mba release tracker lists 224B and 12B active and states no licence or weights status.
What the licence allows is unknown
The card names the Tencent Hy License. The licence file did not load when The Terminal tried it, so whether it permits commercial use, and in which countries, is unconfirmed. The card does not describe training data, beyond saying the base model saw a large amount of synthetic instruction data during mid-training.
The price and the vendor figures
AF.NET reports API pricing of $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached tokens. Full Hy4 costs $0.834 and $2.501, per the same write-up.
- Hy4-flash0.28 USD
- Hy42.5 USD
Source: AF.NET launch write-up, accessed 2026-11-19
The card's own figures are vendor-supplied and unverified by anyone else. It says cost per task is about 1/13 of Hy4 preview and 1/20 of Hy4, and that decoding exceeds 150 tokens per second. The vendor's internal WorkIndex benchmark scores 42.1 against 39.7 for Hy4 preview. Base-model MMLU at 5-shot is 88.7, against 91.6 for the larger Hy4 base.
The per-token prices give a different ratio. On input, $0.070 against $0.834 is about 1/12, and on output, $0.281 against $2.501 is about 1/9. A 1/20 cost per task would need Hy4-flash to spend fewer tokens per task, and the card does not break that out.
Where to use it, and until when
AF.NET lists the model ID hy4-flash on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter planned. Maximum input is 960,000 tokens and maximum output 64,000. WorkBuddy fees are waived through November 30, and TokenHub calls get a 50% discount until 12:00 Beijing time that day.
The model arrives 12 days after the full Hy4 launched on November 4, and while Google's Gemini 4 Neon is ranking third on Artificial Analysis. Google's Gemini 4 Carbon is the other recent launch on the site. Neither has been compared with Hy4-flash by an independent party in the sources reviewed.
The next checkable event is November 30, when the discounts end and the pricing page shows what Hy4-flash costs at list.
Sources
More in A.I.
- 01Anthropic Says Claude Closed a Black Hole Bound Left Between 8.503 and 12The company's post names one model for the result; the paper's own record names two others, and no outside physicist has reviewed it publicly.
- 02Four Students Sue Omnilert and Baltimore County Over the Chip-Bag Gun Alert at Kenwood HighThe Nov. 17 suit targets a $2.6 million school contract, and the district's own timeline shows staff cancelled the alert 1 to 2 minutes after it fired.
- 03Google Ranks Gemini 4 Neon Third on Artificial Analysis; the Site's Own Page Says SixthGoogle's Nov. 18 launch post sells Neon as fast, but Artificial Analysis clocks 21.5 minutes per task against a 12.1-minute median.
- 04Gemini 4 Carbon Takes 36 Minutes a Task; Kernel Counts 7x the Steps of AstraArtificial Analysis puts Carbon at 36.0 minutes per Intelligence Index task against a 12.4-minute median, and a Kernel study traces most of the gap to extra steps.