Product Launch3 min read

Tencent releases Hy4 flash, a low-cost model with a 1M-token context

Tencent's Hy4 flash is a 224B-parameter multimodal model with 12B active parameters, priced at about $0.07 per million input tokens. Tencent says it slightly beats its larger Hy4 Preview on internal work tasks at a fraction of the cost.

AE
Aidose Editorial
Nov 17, 2026

Tencent launched Hy4 flash on November 16, the newest and cheapest model in its Hy family. It has 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active per token, accepts multimodal input and supports a context window of up to 1 million tokens.

Tencent prices the model at about $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens, and calls it "the most cost-effective Hy model to date." The company says the architecture combines hybrid attention, LatentMoE and Engram, giving a KV cache about 1/24 the size of Hy3's and 1/6 the size of Hy4's. Reinforcement learning cut average end-to-end output tokens per task by 50%, and decoding runs at more than 150 tokens per second.

Tencent's performance claims rest on its own internal WorkIndex benchmark, built from typical tasks in its WorkBuddy agent product and scored by human evaluators. On that benchmark, Tencent says Hy4 flash slightly outperforms Hy4 Preview, a 770B-parameter model with 49B active, at roughly 1/13 of its per-task cost, 1/20 of Hy4's and 1/8 of DeepSeek-V4.1-Flash-1017's. In a blind test with 85 internal users across 259 tasks, Tencent reports satisfaction close to Claude Haiku 5.5. The announcement lists public agent benchmarks such as DeepSWE, Terminal-Bench 4.0 and OSWorld 2.1 but does not publish scores for them.

The model is available now through Tencent Cloud TokenHub, where it is 50% off standard pricing at launch, and inside WorkBuddy (free until November 30), Yuanbao, WeCom, QQ, ima and Marvis. Tencent says OpenRouter access is coming soon. Weights are not yet released, though the company says it plans to open-source Hy4 flash.

Tencent positions Hy4 flash for everyday agent work such as deep research, document processing, slide generation, simple front-end pages and computer use, and recommends the full Hy4, released on November 4, for long-horizon software engineering. Developers should treat the cost and quality comparisons as vendor-reported until independent benchmark results are available.

AE
Aidose Editorial
Nov 17, 2026 · 3 min read
Back to News