Two clocks, and a window with no click path
The research post dated 16 November 2026 introduces Hy4 flash with 224B total parameters, including 64B Engram parameters, and 12B active parameters per token. It is multimodal, with a 1M-token context. The price table on that post, per million tokens, is cached input $0.003, input $0.070, output $0.281. The Hy4 row on the same table is $0.042, $0.834, and $2.501. The post says the model is on TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with OpenRouter later. Through 30 November, WorkBuddy access is free, and TokenHub API pricing is 50 percent off. The post does not print an hour for that date, and it does not print a WorkBuddy token cap.
The China price page, updated 2026-11-16 12:15:30, is a different document. Its Guangzhou language-model table prints input 0.5 yuan, output 2 yuan, and a cache hit at 0.02 yuan, per million tokens, for model hy4-flash. Note 1 on that page says calls are settled at 50 percent of the list price until 2026-11-30 12:00:00 Beijing time, then the list price returns. The table does not print a second, discounted row. Note 2 is about Hy4, model id hy4, and that discount runs until 2026-12-31 23:59:59 Beijing time. The price page does not mention WorkBuddy.
| Page | What it prints | Clock |
|---|---|---|
| Research post | WorkBuddy free. API 50 percent off. No token cap. | Through 30 Nov 2026. No hour. |
| China price page, Guangzhou | List: 0.5 / 2 / 0.02 yuan per million tokens. Settled at 50 percent of that list. | Until 12:00:00 Beijing on 30 Nov 2026. |
| Same price page, Hy4 row | List: 6 / 18 / 0.3 yuan. Also 50 percent off. Different model id. | Until 23:59:59 Beijing on 31 Dec 2026. |
| Research USD table | Flash: $0.003 cached, $0.070 in, $0.281 out. Hy4: $0.042 / $0.834 / $2.501. | The 50 percent line has no hour. |
The post's own WorkIndex line says average cost per task is 1/20 of Hy4. A blind test it describes used 85 WorkBuddy users, 259 tasks, and 27 categories, with satisfaction close to Claude Haiku 5.5. Those figures belong to the post. They were not re-measured for this page. The API guide, separate from that test, lists hy4-flash at a 1M context, 960k max input, and 64k max output.
Research post: hy.tencent.ai/research/hy4-flash
Price page: cloud.tencent.com/document/product/1823/130055
Claim the pack, or open WorkBuddy, then send hy4-flash
The new-user pack page says the current activity runs through 31 December 2026 and may be extended or changed. Three claim paths are printed. On the model square, the button is labeled for a new-user free trial: check the models in the dialog, then confirm. On the activation-management page, the card says to claim free resources, then the same dialog. For models that support a free trial, the first call claims the pack automatically.
Under language models, the page says every language model provides 1,000,000 tokens, valid for 1 year from the moment of the claim. Later on the same page, two more lines appear: the quota applies at the main account and is shared by every model on that account, and each model can be claimed once per account. The page does not say which line wins if they disagree. After the free quota, if postpay is off, the service stops. Postpay is switched on from the online-inference page, and tokens are then billed. While the 50 percent window is still open, that bill is half the list price.
Claim steps: cloud.tencent.com/document/product/1823/130053
Claim it: console.cloud.tencent.com/tokenhub
WorkBuddy is the other door, and the docs do not finish it. The research post says access is free in WorkBuddy through 30 November. The TokenHub page about WorkBuddy (product 1823/130673) describes a different setup: install the app, use v4.22.15 or newer, click the custom-model control, pick an enterprise Token Plan preset, and paste an API key. That page does not name hy4-flash and does not mention the 30 November window. If the model switcher already lists Hy4 flash, that matches the post. If it does not, the written path is the API below, not a guessed menu.
Try it: workbuddy.cn/app-download
Try it: workbuddy.ai/app-download
The post also names Yuanbao and does not give a click path.
Try it: yuanbao.tencent.com
Activate first. Then pick the host that matches the account
The international overview says a listed model can be called only after its service is activated in the console. A call before activation returns 402. That error is separate from whether the model supports the protocol. hy4-flash is marked for OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. Auth is Authorization: Bearer and the API key from the TokenHub console.
The international Hy API Guide, last updated 2026-11-16 12:15:11, demonstrates POST https://tokenhub-intl.tencentcloudmaas.com/v1/chat/completions. The sample curl on that page sets model to hy4. The table row for this release sets model to hy4-flash. Swap that one field. Leave the host alone if the account is the international one.
🔍 View full resolution| Account site | Host to paste | Path |
|---|---|---|
| China doc, Guangzhou | https://tokenhub.tencentmaas.com | /v1/chat/completions |
| China doc, Singapore | https://tokenhub-intl.tencentmaas.com | /v1/chat/completions |
| International guide examples | https://tokenhub-intl.tencentcloudmaas.com | /v1/chat/completions |
The China usage page says a call cannot cross region or site. Backup hosts on that page replace .com with .cn. They are fallbacks, not a third region. The international host is the one with cloud in the name. A China account key on the international host is the mismatch this table is for.
Get a key: console.cloud.tencent.com/tokenhub/apikey
China usage page: cloud.tencent.com/document/product/1823/130078
International guide: tencentcloud.com/document/product/1300/80695
Key management, international: tencentcloud.com/document/product/1300/78949
Which door matches the job
Choose WorkBuddy if the date is still inside 30 November 2026 and the model switcher already shows Hy4 flash. The post calls that window free and prints no cap. The install doc does not describe this menu, so a missing name means stop and use the API.
Try it: workbuddy.cn/app-download
Choose the free pack if the call has to come from your own code. The language-model line says 1,000,000 tokens for 1 year. The later lines say the quota is shared on the main account and that each model can be claimed once. Read both before you tick every model in the dialog.
Claim it: console.cloud.tencent.com/tokenhub
Choose the Guangzhou host if the account is on the China site. Paste tokenhub.tencentmaas.com, set model to hy4-flash, and activate the service first. A 402 is the activation check, not a bad key.
Get a key: console.cloud.tencent.com/tokenhub/apikey
Choose the international host if the account follows the English guide. That host is tokenhub-intl.tencentcloudmaas.com. Do not shorten it to the China Singapore host.
Docs: tencentcloud.com/document/product/1300/80695
Choose Hy4, model id hy4, if the job is the long-horizon coding the post assigns to Hy4. On the China price page its 50 percent window ends 31 Dec 2026 at 23:59:59 Beijing time, not on 30 Nov.
Research post: hy.tencent.ai/research/hy4-flash