Models#Benchmark#Model inference
Tencent's Hy4 flash: 224B parameters, 12B active, a tenth of Hy4's API price
Tencent released Hy4 flash on Nov 16: 224B parameters, 12B active, 1M-token context and image input, at 0.5/2 yuan per million tokens, about a tenth of Hy4.

Tencent released Hunyuan Hy4 flash on November 16, the first Flash-tier model in its Hy line. It carries 224B total parameters, 64B of them in an Engram lookup table, activates 12B per token, accepts images and runs a 1M-token context. Tencent Cloud’s TokenHub product page lists it at 0.5 yuan per million input tokens and 2 yuan per million output tokens, with cache hits at 0.02 yuan and a 1M-token free allowance for new accounts. Tencent calls it the best value-for-money Hy model so far and says it plans to open the weights.
Key points
- Specs: 224B total with 64B of Engram parameters, 12B active per token, 1M-token context, image input. Tencent claims decoding above 150 tokens per second without stating the hardware or batch size behind it.
- Cost: on Tencent’s own WorkIndex suite, cost per task is 1/13 of Hy4 preview, 1/20 of Hy4 and 1/8 of DeepSeek-V4.1-Flash-1017 at its max tier and peak-hour rate. Ifanr’s hands-on run points the same way: a 3D interactive web page for 0.23 yuan in half an hour.
- Reach and weights: live on TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter listed as coming. The weights are only “planned” to be open sourced; nothing has been released yet.
Background
Hy4 flash exists because of the full Hy4 release on November 4. That model is the 770B-parameter flagship at 6 yuan input and 18 yuan output per million tokens, with cache hits at 0.3 yuan, and it added vision and computer use to the line. Its predecessor, Hy4 preview, had its weights published under Apache 2.0 back in August. Ifanr noted at the time that Tencent had pre-announced an efficiency-focused flash model alongside Hy4; two weeks later it arrived.
The lane is crowded. DeepSeek’s V4.1-Flash-1017 is 748B total (a 552B backbone plus 196B of Engram), 1M context, MIT licensed, with downloadable weights — Unsloth shipped a quantization last month that fits in 128GB of RAM. Ifanr also notes that during weekday peaks, DeepSeek V4.1 Flash runs at a 0.11x rate multiplier inside WorkBuddy and still queues badly. That is the real contest in the Flash tier: not the price column, but whether a cheap model can actually serve the load. Alibaba’s Qwen4-Max, announced the same day, also led with a 58% drop in cost per task rather than benchmark scores.
The facts
A note on provenance first. The official Hunyuan announcement page renders on the client, so a direct fetch returns only the site name and none of the body. The architecture figures below therefore come from same-day Chinese reporting that relays Tencent’s material, including Cailianshe’s STAR Market Daily and ITHome. The price, the new-account allowance and the distribution channels can be checked directly on the TokenHub product page; the 50% discount and the WorkBuddy quota inside the table come from the relayed announcement, since that page lists only the 1M-token allowance.
| Item | Hy4 flash (as stated by Tencent) |
|---|---|
| Released | 2026-11-16 |
| Total parameters | 224B, including 64B of Engram lookup parameters |
| Active per token | 12B |
| Context | 1M tokens |
| Inputs | Text and images |
| Decoding speed | Above 150 tokens/s |
| KV cache | 1/24 of Hy3, 1/6 of Hy4 |
| API price | 0.5 yuan input, 2 yuan output, 0.02 yuan cache hit per million tokens |
| Promotions | 50% off TokenHub billing until November 30; daily free quota in WorkBuddy from November 16 to 30 |
| Available in | TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, Marvis; OpenRouter planned |
| Weights | Planned to be open sourced, not released |
Three readings need care. The two KV cache multipliers do not share a baseline: 1/24 compares with last generation’s Hy3, 1/6 with the Hy4 flagship, and mixing them changes the story. The 50% discount is a promotion that ends November 30, applied to a list price of 0.5 and 2 yuan; halving the list price gives roughly 0.25 yuan input and 1 yuan output per million tokens, which is our arithmetic and not a published number, so the console bill is the authority — and the flagship Hy4 has a separate 50% discount running to December 31, a different end date. The three cost ratios, 1/20, 1/13 and 1/8, all come from Tencent’s in-house WorkIndex suite, against Hy4, Hy4 preview and DeepSeek-V4.1-Flash-1017 at max tier and peak pricing respectively.
The training and serving changes are relayed as well: mixed attention, LatentMoE and Engram in pretraining, where Engram is an expandable lookup table that trades storage for compute; self-distillation from Hy4 followed by reinforcement learning that cut end-to-end output length by 50%; megakernel, 4-bit quantization and dflash at inference. Tencent’s own routing advice is blunt: deep search, document work, slide decks, simple front ends and computer use go to flash; long-horizon software engineering stays on Hy4.
What others say
The most concrete independent evidence is Ifanr’s hands-on, published at 17:53 on November 16. On the same 3D sailing-ship web page task, Hy4 flash finished in half an hour using about 149,000 tokens plus roughly 4 million cached tokens, which works out to about 0.23 yuan at list prices. DeepSeek V4.1 Flash took an hour and a half, about 300,000 tokens and 16 million cached tokens, for roughly 1.21 yuan. On a deep-research task the two reports came out structurally similar, with DeepSeek spending 10 more minutes, nearly 180,000 tokens and about 0.56 yuan. Across 10 coding and office tasks the verdict was comparable quality with faster first delivery, though the sailing scene had less detail than DeepSeek’s. Ifanr’s conclusion is a split workload: routine high-frequency work to flash, complex codebase edits back to Hy4.
Cailianshe’s STAR Market Daily ran a same-day brief at 13:23 confirming 224B total, 64B of Engram, 12B active, 1M context and the 0.5/2 yuan price, among the earliest Chinese reports. Tencent’s own blind test is a different angle: 85 WorkBuddy users, 259 tasks across 27 categories, satisfaction close to Claude Haiku 5.5 and the shortest task time in the comparison. Its weakness is that every participant is a Tencent product user and the task set is unpublished.
The disagreement is about verification, not about the numbers. “Best value” has no cross-vendor definition, and WorkIndex is a suite Tencent built for its own agent scenarios. Ifanr’s comparison covers one competitor and about ten tasks. At publication time we found no independent leaderboard such as Artificial Analysis re-running Hy4 flash, and the Chinese coverage we could read mostly relays Tencent’s material.
Our take
For teams running dozens or hundreds of agent tasks a day, Hy4 flash belongs in the candidate pool. It combines three things that are rare together in a Chinese closed API: a 1M context, image input and a 12B active footprint, at 0.5/2 yuan. Batch jobs that feed in long documents and screenshots finally have an option priced in loose change. Ifanr’s numbers also make the point that gets missed: in the Flash tier the question is what one finished task costs, not what a million tokens cost, and a cheap model that runs for ninety minutes is not cheap.
The honest costs come in three parts. The weights are not out, so teams that need self-hosting get nothing today — “planned” here is the same wording used for the November 4 flagship, still without a date. The 50% discount ends November 30, after which billing returns to list price. And quality at 1M context, the hardware and batch conditions behind 150 tokens per second, and Engram’s hit rate in long sessions are all unstated in the material available, so they have to be measured rather than assumed.
A rough split: if you need open weights, an MIT license and can live with a 748B deployment, keep looking at DeepSeek-V4.1-Flash-1017; if you need top-end long-horizon engineering, use Hy4; if you are pushing screenshots, PDFs and spreadsheets through an agent and want a fast executor, Hy4 flash is the cheaper tier today. Three things to watch: when OpenRouter goes live, what license and file size the weights ship under, and what the bill looks like on December 1.
How to try it
TokenHub speaks the OpenAI protocol, so the console’s base URL and API key drop into any existing SDK with the model name set to hy4-flash:
curl "$TOKENHUB_BASE_URL/chat/completions" \
-H "Authorization: Bearer $TOKENHUB_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"hy4-flash","messages":[{"role":"user","content":"Explain Engram in one sentence"}]}'
Without code, Hy4 flash is selectable in WorkBuddy, where the daily free quota runs through November 30. The cheapest way to judge it is to run one or two of your own real tasks, note the tokens and wall-clock time, and compare that against whatever you use now.