Skip to content
Tencent Hy4 Flash Slashes Per-Task Cost 20x, Rivals Claude Haiku 5.5

Tencent Hy4 Flash Slashes Per-Task Cost 20x, Rivals Claude Haiku 5.5

Tencent's Hunyuan Hy4 Flash model reduces average per-task cost to 1/20th of Hy4 and 1/13th of Hy4 preview on internal WorkIndex benchmarks, while achieving satisfaction scores near Claude Haiku 5.5 at a fraction of the cost through mixed attention architecture, distillation, RL token reduction, and 4-bit inference acceleration.

AI Frontier Lectures
AI Frontier Lectures
AI Frontier Lectures

On November 16, Tencent Hunyuan released Hy4 Flash, a cost-optimized large language model targeting everyday workflows. The official research page reports that on their internal WorkIndex benchmark, the average cost to complete a task is 1/20 of the base Hy4 model, 1/13 of the larger Hy4 preview, and 1/8 of DeepSeek-V4.1-Flash-1017 at its highest busy-hour pricing tier.

API Pricing (per million tokens)

Pricing is listed in both USD and CNY without conversion:

Hy4 Flash (USD): cache input $0.003, input $0.070, output $0.281

Hy4 (USD): cache input $0.042, input $0.834, output $2.501

Hy4 Flash (CNY): cache input ¥0.02, input ¥0.5, output ¥2

Hy4 (CNY): cache input ¥0.3, input ¥6, output ¥18

Model Architecture

Hy4 Flash totals 224B parameters , of which 64B are Engram lookup tables. Active parameters per token are 120B . The model supports vision and a 1M-token context window . For comparison, Hy4 preview has 770B total and 49B active parameters. Official claims place Flash's performance slightly above the preview tier.

Three Sources of Cost Reduction

The article breaks down cost savings across the model lifecycle:

Pre-training: Mixed attention and latent-variable mixture-of-experts reduce KV cache to 1/24 of Hy3 and 1/6 of Hy4 .

Post-training: Distillation from Hy4 followed by reinforcement learning cuts the number of tokens emitted per task by 50% .

Inference: Large-kernel operators, 4-bit quantization, and a custom accelerator called DFly push decoding throughput beyond 150 tokens/second .

Per-Task Cost Visualization

A bar chart (horizontal axis: per-task cost in CNY) shows:

Hy4: ¥5.7

Hy4 preview: ¥3.2

Hy4 Flash: ¥0.2

Flash is 24.5× cheaper than Hy4 and 13.8× cheaper than preview (the article text cites 20× and 13×; the chart shows 24.5× and 13.8×). Corresponding scores are 51.4 (Hy4), 39.7 (preview), 42.1 (Flash) — Flash scores higher than preview but below the full Hy4. The ¥0.2 figure is per-task cost, not token price.

WorkIndex Blind Evaluation

WorkIndex measures output quality, tokens per turn, and task completion time. The blind test involved 85 internal users , 27 task categories , and 259 tasks . Results (4-point satisfaction scale, median duration, per-task cost in CNY):

Hy4 Flash: 2.79 satisfaction, 6.7 min, ¥0.26

Claude Haiku 5.5 (high tier): 2.81 satisfaction, 6.8 min, ¥4.73

DeepSeek (high tier): 3.09 satisfaction, 10.3 min, ¥1.20

GLM-5.3-Flash (high tier): 2.53 satisfaction, 7.6 min, ¥0.81

Flash matches Haiku 5.5 on satisfaction and latency while costing ~18× less. The ¥0.26 per-task cost here differs slightly from the chart's ¥0.2.

Availability & Roadmap

Hy4 Flash is accessible via Tencent Cloud TokenHub, WorkBuddy, Yuanbao, Enterprise WeChat, QQ, ima, and Marvis; OpenRouter integration is underway. Until November 30, WorkBuddy is free and TokenHub API calls are 50% off. The team recommends Flash for daily tasks and retains Hy4 for long-horizon software engineering. Model weights are planned for release but not yet available.

Sources: Hunyuan research page ( https://hy.tencent.ai/research/hy4-flash) and Tencent Chinese announcement (

https://www.tencent.com/zh-cn/tencent-launches-hy4-flash-to-power-high-speed-cost-efficient-ai-deployment-for-everyday-workflows/

).

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 accelerationreinforcement learningmodel distillationTencent HunyuanLLM cost optimizationClaude Haiku 5.5Hy4 FlashWorkIndex benchmark
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.