Reportedly: Tencent releases Hy4 flash: 224B total parameters, 12B active, per-task cost about 5% of Hy4
Reported by IT Home and ifanr · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Tencent's Hunyuan team released Hy4 flash, a 224B-parameter model with 12B active per token, 1M-token context, and multimodal input. API pricing is much lower than Hy4; on WorkIndex, per-task cost is 1/20 of Hy4 and 1/8 of DeepSeek-V4.1-Flash-1017, IT Home reported.
China context
- Original name
- 腾讯混元 Hy4 flash
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- If you build on Tencent Hunyuan APIs, Hy4 flash can cut per-task costs to ~5% of Hy4 for similar capability, but verify performance on your own tasks before migrating.
- For investors
- Tencent's aggressive price-performance move intensifies the Chinese model price war; watch for margin pressure on smaller labs and whether enterprise adoption accelerates.
Translated from Chinese. Quotes and facts link to the original sources.
Tencent's Hunyuan team announced Hy4 flash, positioned as the most cost-effective Hy model to date. It has 224B total parameters including 64B Engram parameters, activates 12B per token, supports multimodal input and 1M-token context. KV cache is reduced to 1/24 of Hy3 and 1/6 of Hy4. Post-training uses self-distillation from Hy4 and RL to compress output length by 50%. Inference uses megakernel, 4-bit quantization, and dflash, achieving over 150 tps decoding. On the WorkBuddy WorkIndex benchmark, per-task cost is 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (max tier, busy-hour price).
Hy4 flash uses hybrid attention, LatentMoE, and Engram (a scalable lookup-table parameter module) to expand capacity at low compute cost. KV cache is 1/24 of Hy3 and 1/6 of Hy4. Post-training self-distills from Hy4 and uses RL to halve output length. Inference combines megakernel, 4-bit quantization, and dflash for >150 tps decoding.
Developers using Tencent Hunyuan APIs can now get Hy4-level intelligence at about 5% of the per-task cost, directly undercutting DeepSeek's flash-tier pricing by 8x on the cited benchmark. This pressures rival Chinese model providers to match the price-performance or lose cost-sensitive enterprise workloads.
For enterprises running high-volume agentic or multimodal workloads, Hy4 flash offers a path to cut inference costs by up to 95% versus Hy4 while retaining comparable capability, assuming the benchmark transfers to production tasks.
Watch for independent verification of the WorkIndex cost claims and real-world latency at 1M context. A key signal is whether Tencent extends the same Engram/LatentMoE architecture to larger Hy models or open-weights releases.