WorkBuddy Drops Hy4 Flash: 224B MoE, 1M Context, Free for a Limited Time
Title: WorkBuddy has launched a new free model (Hy4 flash), let's take advantage of it.
Author: SimonKing
Tags: backend, Java, programmer
Follow my WeChat official account: 【编程朝花夕拾】 for first-hand content.
01 Introduction
I had been using Hy3 and Hy4 for some small tasks, but today when I used them, I found that Hy4 started charging.
Hy3 also disappeared from the model list. So I gave up, but in the afternoon a popup appeared, surprising me: a new free model Hy4 flash was released, and it can be used again.
02 Introduction
What exactly is Hy4 flash? In one sentence, it's the most cost-effective Hy model from Hunyuan so far, a lightweight main model designed for high-frequency, large-scale daily productivity tasks.
Specs overview:
| Metric | Hy4 flash | Reference comparison |
|---|---|---|
| Total parameters | 224B (including 64B Engram parameters) | Hy4 preview is 770B |
| Active parameters | 12B | Hy4 preview is 49B |
| Context length | 1M | Friendly for long documents/large repositories |
| Multimodal | Native support (visual understanding) | Can read images, screenshots |
| Decoding speed | Over 150 tps | Extremely optimized inference stage |
| API price | Input ¥0.5 / Output ¥2 per million tokens | Hy4 is ¥6 / ¥18 |
Core advantages:
- Cost breaks through the floor: Input 0.5 yuan per million tokens, output 2 yuan, and only 0.02 yuan when hitting the cache, an order of magnitude cheaper than
Hy4. - Extremely simplified architecture: Mixed attention +
LatentMoE+Engramtriad, withKV cacheonly 1/24 of Hy3 and 1/6 of Hy4, cutting memory pressure in half and then again. - Effectiveness improves rather than decreases: Through self-distillation and
Hy4 +reinforcement learning, overall performance is slightly better than the much largerHy4 preview. - Real-task oriented: Evaluation doesn't rely on public benchmark scores far from daily work, but uses
WorkBuddyreal scenarios as the evaluation set.
The official launch list is also impressive: Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, Marvis have all been launched simultaneously, and OpenRouter integration is coming soon; Tencent Cloud ADP also supports it, so you can directly build enterprise-level Agents with it.
By the way: Hy4's free access has ended, and Hy3 has been removed from the app. Now the free quota is all given to Hy4 flash, with a fixed daily allowance. After hitting the limit, you can continue using it by consuming points, and the page shows the reset time; there may be queues during peak hours.
03 Trial
Opening the client, it works.
Let's try it and see the results.
Using Hy4 flash to optimize my previous notes project, it feels similar to Hy4, basically one-shot optimization works.
04 Surprise
I checked the WorkBuddy message center and found that the Hy4 flash announcement was pushed on 2026.11.16 15:44, and the free period lasts until the end of the month.
I also discovered that the points I had used for the deepseek 4.1 Flash model were returned because of a discount adjustment.
Check the bill to see if it's real.
The validity period is actually one year, not bad!
05 Summary
No matter the model, there's a free model with decent speed, so why not try it. Interested friends can go give it a shot.