跪拜 Guibai
← All articles
Backend · Java · Programmer

WorkBuddy Drops Hy4 Flash: 224B MoE, 1M Context, Free for a Limited Time

By SimonKing ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

At 0.5 yuan per million tokens, Hy4 Flash undercuts most Western API pricing by an order of magnitude, and its 1M context and vision make it a viable replacement for heavier models on routine tasks. The free tier on WorkBuddy removes the subscription barrier, but daily caps and peak-hour queues are the trade-offs.

Summary

Hy4 Flash arrives on WorkBuddy as a free tier model built by Hunyuan. It combines a 224B-parameter mixture-of-experts with 12B active parameters, uses a mixed-attention plus LatentMoE plus Engram architecture, and supports a 1 million token context with native vision. Decoding runs over 150 tokens per second.

The API pricing goes from 0.5 yuan per million input tokens to 2 yuan per million output tokens, with cache hits at 0.02 yuan. That's an order of magnitude cheaper than the now-paid Hy4, and the developer behind the report notes the free quota has shifted to Hy4 Flash, with daily allotments and points as a fallback.

Hy4 Flash beats the larger 770B Hy4 preview on overall effect, thanks to self-distillation and reinforcement learning. Hunyuan evaluates it on real WorkBuddy tasks rather than public benchmarks. The model is already live across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with OpenRouter on the way. The free period runs until the end of the month.

Takeaways
— 224B total parameters, 12B active, including 64B Engram parameters
— 1M token context length, native vision support
— Decoding speed above 150 tokens per second
— API price: 0.5 yuan per million input tokens, 2 yuan per million output tokens; 0.02 yuan on cache hit
— KV cache is 1/24 of Hy3 and 1/6 of Hy4
— Performs slightly better than Hy4 preview (770B) via self-distillation and reinforcement learning
— Evaluated on real WorkBuddy tasks, not public benchmarks
— Available on TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, Marvis; OpenRouter planned
— Hy4 free access ended; Hy3 removed
— Daily free quota with points as fallback; queues during peak
— Free until end of month, announced 2026.11.16
— Points spent on DeepSeek 4.1 Flash refunded due to discount adjustment
Conclusions

Pricing cache hits at 0.02 yuan per million tokens makes repeated calls effectively free, encouraging high-frequency use in production loops.

A 12B-active model outperforming a 770B preview shows that architecture and training quality can trump raw parameter count.

Hunyuan's decision to evaluate on real WorkBuddy scenarios rather than public benchmarks prioritizes practical utility but makes cross-model comparisons difficult.

The refund of points for DeepSeek 4.1 Flash indicates active price competition among Chinese LLM providers, with adjustments flowing back to users.

Concepts & terms
LatentMoE
A mixture-of-experts routing mechanism that activates a small subset of expert networks per token based on a learned latent representation, reducing computational cost while keeping model capacity high.
Engram
A component in Hunyuan's architecture that stores and retrieves encoded memory (the 64B Engram parameters), providing a compact long-term memory for the model, likely improving context retention and efficiency.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗