Tencent Hunyuan has released Hy4 three times in the past three months, and each time the attitude is different.
The Hy4 preview in August was released as open source under the Apache 2.0 protocol. On that day, four hosting sites were listed, and the deployment scripts of vLLM and SGLang were provided together. The official version of Hy4 was launched on November 4, with an input cost of 6 yuan, an output of 18 yuan, and a cache hit of 0.3 yuan per million tokens. Artificial Analysis gave it an intelligence index of 34.6. On November 16, that is, yesterday, Hy4 flash came: input of 0.5 yuan, output of 2 yuan, and cache of 0.02 yuan.
It took 12 days for the price to drop from 18 yuan to 2 yuan.
But there is a change in the opposite direction on the same timeline: the preview can be downloaded on the day it is released, and flash only has the sentence "planned to be open source", with no date, no agreement, no model card, and no technical report.
So this comparison is a bit special. Hunyuan Hy4 flash is competing for the cheap price, and its head-on competitors are DeepSeek V4.1 Flash 1017 and Zhipu’s GLM-5.3-Flash. The latter two are both under the MIT protocol and hold the weight; the new one has the lowest price, but except for the chart drawn by Tencent itself, there is no third-party results for the time being.
The two old rivals have each changed their rules.
DeepSeek V4.1 Flash was released on September 10, under the MIT license. On October 17, the new snapshot 1017 directly replaced the original version, and the API name did not change a word. The people who adjusted this interface in those few days already got another model, but they couldn't see it in the code.
Its size is 552B total parameters, 16B activations, 1 million contexts, a maximum output of 384,000 tokens at a time, and a weight of about 510GB. AA intelligence index is 32.6 (max level), Terminal-Bench 2.1 gets 90.6, and DeepSWE v1.1 is 74.2.
GLM-5.3-Flash was released on August 26, open source on the same day, under the MIT protocol, with 320B total parameters, 18B activations, the same 1 million contexts, and a maximum output of 128,000. It was offered at a 50% discount when it first launched, and it expired on September 9th without renewal. AA intelligence index is 27.8, the lowest among the three; AA programming index is 71.5.
The specifications of the three companies are put together:
| Hy4 flash | DeepSeek V4.1 Flash 1017 | GLM-5.3-Flash | |
|---|---|---|---|
| General parameters / activation | 224B / 12B | 552B / 16B | 320B / 18B |
| Input (USD/million) | 0.070 | 0.15 free time / 0.30 busy time | 0.15 |
| Output (USD/million) | 0.281 | 0.60 free hours / 1.20 busy hours | 0.50 |
| cache hit | 0.003 | 0.003 free time / 0.006 busy time | 0.03 |
| weight | Not released | MIT | MIT |
| AA Intelligence Index | 无 | 32.6 | 27.8 |
The US dollar price of Hy4 flash is converted based on the domestic prices of 0.5 yuan, 2 yuan, and 0.02 yuan. Tencent does not announce the overseas price separately.
At what hour is "one-eighth" calculated?
The loudest sentence in Tencent’s propaganda: the cost of a single task is one-eighth that of DeepSeek V4.1 Flash 1017.
There are two conditions behind this sentence: DeepSeek turns on the max inference file and charges according to the busy hour price.
DeepSeek's busy hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which in Beijing time is 9 a.m. to 12 a.m. and 2 p.m. to 6 p.m. Most of the work that domestic companies do during the day falls right into this. This premise itself is not biased. The problem is that it is only one of three algorithms.
In terms of pure unit price ratio: DeepSeek input is 4.3 times that of Hy4 flash during busy times, and output is 4.3 times; both are only about 2.1 times during idle times. The remaining gap comes from the shorter Hy4 flash output. Tencent said that it uses Hy4 self-distillation to enhance reinforcement learning and cut the task output length in half. The money saved depends on whether your task can also be shortened.
Take a common agent request and calculate it: 200,000 tokens are input, 90% of the cache hits, and 5,000 tokens are output.
Hy4 flash is about $0.0034. DeepSeek is $0.0065 when idle and $0.0131 when busy. GLM-5.3-Flash is $0.0109.
The difference is 2x to 3.9x, not 8x.
This account also contains GLM’s pitfalls. Its output unit price is $0.50, which seems cheaper than DeepSeek's idle price of 0.60; the cacheable hit price is $0.03, 10 times that of the other two. If the tasks of long context and high cache hits are left to it, the ledger will be more than 60% more expensive than DeepSeek when it is idle. People who only look at the input and output columns for price selection can easily make mistakes here.
Tencent’s set of internal tests should be read separately when reading.
The results obtained by Hy4 flash all come from Tencent’s own business and evaluation sets.
The blind test was conducted on 85 WorkBuddy users, 27 types of scenarios, and 259 daily tasks. The conclusion was that the satisfaction level was "close to" Claude Haiku 5.5, with the lowest cost and shortest time-consuming per task. The participants are internal employees of Tencent and use Tencent’s own products. "Close" does not mean "exceed" either.
There are several items in the business data that are quite specific: WeChat Pay's static scan takes 37% less time and 46% fewer tokens at the same detection quality; King of Honor's failure root cause analysis shows that the number of requests is 22% to 39% less than DeepSeek and GLM, and the number of tokens is 40.5% to 61.2% less.
There is another line in the same set of data, but the promotional release didn’t mention it much.
The number of tool calls, Hy4 flash is 51.2% more than DeepSeek and 7.7% more than GLM. Tencent’s explanation is “more and shorter round trips.” If your tool is backed by a pay-per-use search API, or each call has to wait for a slow interface, part of the token money saved may be eaten up by the number of calls.
The other two are shortcomings stated by Tencent itself: long-range coding tasks are slightly lower than those of Hy4, and the official recommendation for complex software engineering is to leave it to Hy4; its positioning is to do office work such as PPT, document modification, table modification, and in-depth search. DeepSeek V4.1 Flash scored 90.6 on Terminal-Bench 2.1, and Hy4 flash currently cannot provide comparable figures. Its results on DeepSWE and Terminal-Bench 4.0 only appear on Tencent's page in the form of pictures.
The problem with DeepSeek is not the API
There are two main issues with V4.1 Flash 1017.
One is the snapshot replacement mentioned earlier. It was changed quietly on October 17, the interface name remained unchanged, and the evaluation baseline drifted accordingly. Teams that have done regression testing in the production environment will not match the scores that week.
Second, the shortage of computing power has spread to the product side. In the past two days, some users have received prompts that long conversations that have not been used in web pages and apps for more than 7 days cannot be continued. A single input is stuck at about 40,000 words. If it is used too much, it will directly prompt that the quota is full. The same limitations don't apply to the API, but doubling the price during busy times is itself the flip side of the same thing: not enough cards during the day.
Where did GLM-5.3-Flash lose?
Cache price aside, its Intelligence Index of 27.8 is nearly 5 points lower than DeepSeek. The first word delay measured on the third-party routing platform is 3.82 seconds for p50 and 10 seconds for p95, which is not very friendly to dialogue products that require immediate feedback.
Its benefits are also real: the weights are on Hugging Face, both FP8 and BF16, and it natively supports image and video input. For teams that need privatization and don’t want to carry the weight of 510GB, it is much easier to implement than DeepSeek.
Should we cut off the traffic?
My judgment is not to cut the whole thing now.
Hy4 flash has two dates to keep an eye on: November 30 when WorkBuddy is free and TokenHub is 50% off; and when it will actually release its weight. Tencent has said that it will connect to OpenRouter, but several third-party routers have inconsistent opinions on whether it has been put on the shelves.
For office, document, and short-output batch tasks, you can use part of the traffic to try it out on TokenHub first, and record the number of tool calls in your account.
Programming and long-link agents stay in DeepSeek V4.1 Flash, and the batch processing that can be moved is moved to Beijing time evenings and weekends, and the unit price is directly discounted in half.
If you want to deploy it privately or the data cannot leave the intranet, GLM-5.3-Flash is the only realistic choice for Sanjia. The premise is that you don’t let it run high-caching long context tasks.
Site entrance: Tencent Hunyuan Large Model, DeepSeek, Zhipu AI, Tencent Yuanbao.
Related comparisons: "MiniMax M3.1 Flash and DeepSeek V4.1 Flash 1017: The price is exactly the same during leisure time, DeepSeek is twice as expensive during the day on weekdays", "Gemini 4 Hydrogen and DeepSeek V4.1 Flash 1017: The difference is five times or fifty times, it depends on whether the 100,000 tokens hit the cache", "The 2.99 points of Hunyuan Hy4 preview does not mean anything, GLM-5.3 and Kimi K3’s account should be settled》