Tencent released Hy4 Flash on November 16, 2026. It is the cheap, fast model in the Hy4 family. Tencent’s launch post leads with a blind test where Hy4 Flash scored close to Claude Haiku 5.5 for a fraction of the cost. That is true. The same chart shows a different model in first place: DeepSeek V4.1 Flash 1017.
Model ID. Tencent API:
hy4-flash. OpenRouter: not listed yet. Weights: not released.
What is Hy4 Flash?
Hy4 Flash is a mixture-of-experts model that Tencent distilled from its flagship Hy4 and then trained further with reinforcement learning. According to Tencent’s announcement, the RL step cut output length per task by 50%.
| Spec | Hy4 Flash |
|---|---|
| Total parameters | 224B |
| Of which Engram | 64B |
| Active per token | 12B |
| Context | 1M tokens |
| Input | Text and images |
| Output speed | Over 150 tokens/s |
| Architecture | Hybrid attention, LatentMoE, Engram |
| Weights | Not released |
Tencent-reported, from the Hy4 Flash announcement.
Tencent says the KV cache is 24 times smaller than Hy3’s and 6 times smaller than Hy4’s. It serves the model with 4-bit quantization. Tencent also says what the model is for: search, documents, slides, simple frontend work, and computer use. For long software projects it recommends the full Hy4.
The blind test, and who won it
Tencent asked 85 internal WorkBuddy users to rate answers on 259 real tasks of 27 types, without knowing which model wrote them. The rivals ran at their high setting. Here is the full table from the announcement.
| Model | Satisfaction (out of 4) | Median time | Cost per task |
|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | 3.09 | 10.3 min | ¥1.20 |
| Claude Haiku 5.5 | 2.81 | 6.8 min | ¥4.73 |
| Hy4 Flash | 2.79 | 6.7 min | ¥0.26 |
| GLM-5.3-Flash | 2.53 | 7.6 min | ¥0.81 |
Tencent-reported, from the blind test chart in the Hy4 Flash announcement. In its appendix notes, Tencent says it computes costs from each model’s official API price, converts dollar prices at 7.2 yuan per dollar, and prices DeepSeek at its peak-hour rate. Off-peak, DeepSeek costs half.
DeepSeek V4.1 Flash 1017 won on satisfaction, 0.28 points ahead of Haiku and 0.30 ahead of Hy4 Flash. Most launch coverage leaves this out.
Hy4 Flash and Haiku 5.5 are tied in practice, 2.79 against 2.81, and finished in almost the same time, 6.7 against 6.8 minutes. Hy4 Flash was the fastest by 0.1 minutes.
The cost gap is large. Hy4 Flash cost ¥0.26 per task, Haiku ¥4.73, and DeepSeek ¥1.20 at its peak rate.
This is Tencent’s own test, with Tencent’s users, on tasks Tencent picked. Treat it as a first signal until someone else runs one. We tested DeepSeek V4.1 Flash 1017 and Haiku 5.5 on a coding task last month; see our DeepSeek V4.1 Flash 1017 test for those results.
WorkIndex: Tencent’s office-task score
Tencent also reports a composite score called WorkIndex, the average of 8 internal office tasks such as slide generation, Word and Excel fixes, and audio and video editing.
| Model | WorkIndex | Cost per task |
|---|---|---|
| Hy4 | 51.4 | ¥5.7 |
| DeepSeek V4.1 Flash 1017 | 44.4 | about ¥1.64 |
| Hy4 Flash | 42.1 | ¥0.2 |
| Hy4-preview | 39.7 | ¥3.2 |
Tencent-reported, from the charts and appendix table in the Hy4 Flash announcement. The DeepSeek row is the average of its eight WorkIndex scores in the appendix, at the peak-hour price.
Hy4 Flash gives up about 9 points against Hy4 and costs about 24.5 times less per task, by Tencent’s chart. Tencent’s text rounds this to 1/20. DeepSeek V4.1 Flash 1017 scores 2.3 points above Hy4 Flash on WorkIndex as well.
Coding and agent benchmarks
Tencent’s appendix also covers public agent benchmarks. Tencent ran all models itself, at the highest reasoning setting each offers.
| Benchmark | Hy4 Flash | Hy4 | Hy4-preview | DeepSeek V4.1 Flash 1017 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 19.2 | 44.4 | 26.3 | 40.0 |
| ProgramBench (almost solved) | 16.0 | 30.0 | 17.5 | 29.0 |
| DeepSWE v1.1 | 64.6 | 74.9 | 64.3 | 73.8 |
| OSWorld 2.1 (partial score) | 68.1 | 78.9 | n/a | 68.4 |
| HLE-Full (with tools) | 55.2 | 59.1 | n/a | 57.1 |
Tencent-reported, from the appendix table in the Hy4 Flash announcement. Terminal-Bench 4.0 ran in the Claude Code harness. The ProgramBench figure is the “almost solved” score, not full solves.
Terminal-Bench 4.0 is the closest of these to how a coding agent works. There Hy4 Flash scores 19.2, less than half of Hy4 and of DeepSeek V4.1 Flash 1017 in Tencent’s own run with the same harness, and below the older Hy4-preview. This matches Tencent’s advice to use the full Hy4 for long software work.
How much does Hy4 Flash cost?
| Rate per 1M tokens | Hy4 Flash list | Hy4 Flash until Nov 30 | Hy4 list |
|---|---|---|---|
| Input | ¥0.5 | ¥0.25 | ¥6 |
| Cached input | ¥0.02 | ¥0.01 | ¥0.3 |
| Output | ¥2 | ¥1 | ¥18 |
From Tencent Cloud’s TokenHub pricing page, November 16, 2026.
TechNode converts the Flash list rates to about $0.07 input and $0.28 output. TokenHub charges half until November 30, 12:00 Beijing time, and the price does not change with context length. Hy4 is also 50% off, until December 31.
Hy4 Flash is free in the WorkBuddy app until November 30, according to Tencent. The API has no free quota. We did not find an official international dollar price.
Hy4 Flash vs Hy4 on OpenRouter
The full Hy4 is already on OpenRouter as tencent/hy4. Its endpoint listing shows one provider, Tencent, at $0.417 input and $1.25 output per million tokens, with a 50% discount flag and fp4 quantization. Expect that price to double when the discount ends. It supports tool calls and a reasoning effort setting, and it has a 1M context with up to 65,536 output tokens.
Hy4 scores 35 on the Artificial Analysis Intelligence Index, below the median of 41. Artificial Analysis and Vals have not published results for Hy4 Flash yet.
Can you run Hy4 Flash locally?
Not yet. Tencent has not released the weights, and we found no Hy4 Flash repository on Hugging Face, ModelScope, or GitHub. Tencent says it plans to open them, with no date. llama.cpp supports Hy4-preview, but not the new hybrid attention, LatentMoE, and Engram layers in Flash.
When weights arrive, here is our rough estimate, not a measurement. At about 4.8 bits per weight, a 4-bit build of 224B parameters needs around 130 GB. On a 128 GB Mac that does not leave room, so a 3-bit build of about 95 to 100 GB is the realistic option. The small KV cache helps with long context.
How to try Hy4 Flash today
Today the only way in is Tencent’s own platforms: free in the WorkBuddy app, or at half price through the TokenHub API, both until November 30.
We have not tested Hy4 Flash in Atomic Agent. When it appears on OpenRouter, we will run it on the same coding task as our DeepSeek V4.1 Flash 1017 test and add the results here, with cost per task and time.
Hy4 is available through OpenRouter today. To use it, download Atomic Agent for Mac, Windows or Linux, or install the terminal version. It is open source and works with your own API keys.
curl -fsSL https://atomicagent.io/install | shatomic-agentConnect OpenRouter with your own key and pick tencent/hy4 from the model list; see our OpenRouter guide. If you want a cheap Flash model for coding today, see DeepSeek V4.1 Flash vs MiniMax M3.1 Flash, and for agents rather than models, our list of open-source Claude Code alternatives.
The takeaway
By Tencent’s own numbers, Hy4 Flash matches Claude Haiku 5.5 on satisfaction at about one eighteenth of the cost per task in Tencent’s test. DeepSeek V4.1 Flash 1017 scored higher in the same test and cost about 4.6 times more per task than Hy4 Flash at its peak rate. On Terminal-Bench 4.0 in Tencent’s own runs, Hy4 Flash scored 19.2, against 44.4 for Hy4 and 40.0 for DeepSeek V4.1 Flash 1017.
For documents, search, and office tasks, it is worth a free try in WorkBuddy, or half price on TokenHub, before November 30. For coding agents, wait for independent results and for the OpenRouter listing. We will publish our test when it is listed.


