Skip to content
VECTOR WIREAI INTELLIGENCE
UTC
Home

Tencent cuts Hy4 per-task cost 95% with new flash-tier model

224B total parameters, 12B activated per pass, 1M-token context window.

Tencent released Hy4 flash on November 16, its first flash-tier language model, priced at roughly one-twentieth the per-task cost of the full Hy4 on the company's internal WorkIndex benchmark2,3.

The model has 224 billion total parameters, including 64 billion Engram parameters, but activates only 12 billion per forward pass1. It accepts multimodal input and supports a context window of up to one million tokens.

Pricing and availability

API list prices are $0.07 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens. In yuan terms, that is 0.5 yuan input, 2 yuan output and 0.02 yuan on cache hits, compared with 6 yuan, 18 yuan and 0.3 yuan for the full Hy4. Through November 30, Hy4 flash is free in WorkBuddy and available at half the list price on Tencent Cloud's TokenHub API.

The model is live in WorkBuddy, Yuanbao and TokenHub. Tencent plans to release the model's weights and make it available on OpenRouter.

Architecture and inference stack

Tencent attributes the cost reduction to changes across pre-training, post-training and inference. Pre-training introduces hybrid attention, LatentMoE and Engram, a lookup-style parameter module. The KV cache is 1/24 the size of Hy3's and 1/6 of Hy4's. Post-training distills from Hy4 and applies reinforcement learning to cut end-to-end task output length by 50%. At inference, megakernels, 4-bit quantization and a system called DFly decode at more than 150 tokens per second.

Benchmark claims and limits

Tencent's performance numbers come from WorkIndex, an internal benchmark built from typical WorkBuddy tasks such as slide editing, document work, spreadsheet repair, audio/video processing and visual understanding, combined with human ratings; WorkIndex is not public. In a blind test in which 85 WorkBuddy users rated outputs across 259 tasks, Tencent said Hy4 flash scored close to Claude Haiku 5.5 in user satisfaction and had the lowest cost and shortest completion time among the models tested. On collaborative PPT editing, Hy4 flash scored 74.0 versus Hy4's 78.9, but per-task cost fell from 1.714 yuan to 0.105 yuan. The model trails notably on complex tasks such as Terminal-Bench 4.0. Tencent said it focuses on everyday work rather than public leaderboard scores.

The release follows Tencent's November 5 expansion of Hy4 inference capacity tenfold after day-one demand on WorkBuddy overwhelmed quotas[3].

ANALYSIS Shipping a flash tier less than two weeks after that capacity crunch gives Tencent a far cheaper model to absorb high-volume WorkBuddy traffic without proportional infrastructure cost.

Was this useful?

The day’s vector, in your inbox.

Vector Wire Daily: one dispatch each publishing day.