Skip to content
Tencent Hy4 Flash: 12B Active Params, 1/8 DeepSeek Cost, Matches Claude Haiku 5.5

Tencent Hy4 Flash: 12B Active Params, 1/8 DeepSeek Cost, Matches Claude Haiku 5.5

Tencent's Hy4 Flash, a 224B MoE model with 12B active parameters, achieves 150+ tokens/s decoding speed at 1/20 the cost of its predecessor, outperforms DeepSeek-V4.1-Flash on cost efficiency, and matches Claude Haiku 5.5 user satisfaction in blind tests across 259 office tasks.

SmartApe AI
SmartApe AI
SmartApe AI

Tencent has released Hy4 Flash, a new mixture-of-experts model in its Hunyuan series, targeting high-throughput, low-cost inference for everyday office workloads.

Model Architecture and Specifications

Hy4 Flash totals 224 billion parameters, including 64 billion Engram parameters, but only 12 billion parameters are activated per forward pass. It supports multimodal input and a 1-million-token context window. The model achieves a decoding speed exceeding 150 tokens per second.

Cost Efficiency Claims

According to Tencent's published data, Hy4 Flash reduces per-task cost to 1/13 of Hy4 Preview and 1/20 of the original Hy4 on the internal WorkIndex benchmark. Compared to DeepSeek-V4.1-Flash-1017's peak-hour max tier pricing, Hy4 Flash completes similar tasks at roughly 1/8 the cost, prompting community remarks that it is "cheaper than Liang Wenfeng's model."

Technical Optimizations

The cost reduction stems from optimizations across the full stack:

Pre-training: Hybrid attention, LatentMoE, and Engram mechanisms preserve knowledge capacity while controlling compute.

KV Cache: Cache size is reduced to 1/24 of Hy3 and 1/6 of Hy4, lowering memory overhead for long contexts.

Post-training: Self-distillation from the larger Hy4 model transfers capabilities; reinforcement learning cuts end-to-end task output length by 50%.

Inference Deployment: Megakernel, 4-bit quantization, and DFly technologies further accelerate execution, enabling the 150+ tokens/s decode rate.

Evaluation Methodology

Beyond public benchmarks, Tencent built WorkIndex, an internal suite covering eight office-task categories: PPT creation, PPT editing, PPT human-AI co-writing, Markdown co-writing, document editing, spreadsheet repair, audio/video processing, and visual understanding. Additionally, seven public agent benchmarks were used: DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE, and OSWorld 2.1.

Hy4 flash Bench图表-全.png
Hy4 flash Bench图表-全.png

Results show Hy4 Flash slightly outperforms Hy4 Preview overall while drastically cutting running cost.

Blind User Study

To validate real-world utility, Tencent conducted a blind test with 85 internal WorkBuddy users across 27 task categories totaling 259 daily work tasks. The study measured delivery quality, task latency, and user satisfaction.

Blind test results
Blind test results

Hy4 Flash achieved user satisfaction nearing that of Claude Haiku 5.5, while recording the lowest per-task cost and shortest completion time among all models compared.

Deployment and Usage Recommendations

Hy4 Flash is already integrated into multiple Tencent AI products; WorkBuddy users can invoke it for document processing, PPT generation, information retrieval, and similar tasks. Tencent provides explicit routing guidance:

Prefer Hy4 Flash for deep search, document handling, PPT generation, simple frontend development, and Computer Use scenarios.

Reserve the more capable Hy4 for long-horizon software engineering tasks requiring complex reasoning, continuous planning, and multi-step execution.

Open-Source Roadmap

Tencent has announced plans to open-source Hy4 Flash and commits to high-frequency iterative releases thereafter.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

Inference OptimizationMixture of ExpertsOpen Source LLMLLM EvaluationHunyuanTencent AIHy4 Flash
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.