Tencent releases Hy4 Flash, an open-weight model it says does everyday tasks at a twentieth of Hy4's cost

AI & Compute

Tencent Open-Sources Hy4 Flash, Claiming Tasks at 1/20th of Hy4's Cost

Tencent released Hy4 Flash as open weights on Nov. 17, claiming a fifth of Hy4's quality gap at a twentieth of its cost, under a licence with revenue thresholds for API sellers.

By
Nathan Brooks
Filed
Channel
AI & Compute
Read
3 min read

Tencent published Hy4 Flash, Hy4 Flash-FP8 and a pre-trained base model as open weights on Nov. 17, claiming its cheapest Hy model to date at one twentieth of Hy4's cost per task. The repositories appeared on Hugging Face, ModelScope and GitCode one day before the company reported third-quarter results in which costs for new AI products weighed on profit. The timing was, presumably, a coincidence.

Hy4 Flash is a mixture-of-experts design with 220 billion total parameters, of which 64 billion are n-gram embedding parameters that Tencent calls Engram. The model activates roughly 12 billion parameters per token, accepts text and images, and supports a context window of up to 1,048,576 tokens. Tencent positions it for deep search, document processing, slide creation, simple front-end demos and computer use. Long-horizon software engineering remains the job of the larger Hy4.

What does Tencent claim on cost and quality?

Every figure in this section comes from Tencent itself. On WorkIndex, an internal benchmark of eight task sets built around its WorkBuddy office agent — slides, documents, spreadsheet repair, audio and video processing, visual understanding — Hy4 Flash scores 42.1. Hy4 preview scores 39.7 and full Hy4 scores 51.4.

The cost claims are the pitch. Average cost per task is 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 at that model's peak-hour pricing, according to the company.

A blind test adds a human datapoint. Eighty-five internal WorkBuddy users ran 259 tasks; Hy4 Flash scored 2.79 out of 4 for satisfaction against 2.81 for Anthropic's Claude Haiku 5.5. It cost RMB0.26 per task with a median completion time of 6.7 minutes — the lowest cost and shortest time among the models tested, Tencent says.

On few-shot tests of the base model, Tencent reports 88.7 on MMLU and 71.1 on MMLU-Pro, versus 89.7 and 68.6 for DeepSeek-V4-Flash-Base in its own runs.

How is the model built?

Tencent attributes the savings to three changes from Hy4:

  • Interleaved sparse and sliding-window attention at a 1:3 ratio.
  • Routed experts working in a smaller latent space: each token uses 16 of 512 experts plus one shared expert.
  • Engram bigram and trigram embeddings.

Together, these cut the memory cache required for long inputs to 1/24 of Hy3's and 1/6 of Hy4's. Reinforcement learning trimmed average output length by 50 percent, the company says. A built-in 0.85 billion-parameter drafter for speculative decoding, combined with 4-bit quantization, pushes decoding past 150 tokens per second. The model runs on vLLM and SGLang today; support in Hugging Face's Transformers library is pending.

What does the licence restrict?

The weights ship under a Tencent Hy licence, not a standard open-source licence. A company selling general-purpose API access must obtain a separate written licence if its worldwide revenue exceeds $50 million over 12 months, or before deploying the model in a product with more than 100 million monthly active users. Internal use is exempt. The threshold structure mirrors the approach of other Chinese labs: open enough to seed adoption, gated enough to monetize the largest resellers.

Context matters here. Tencent's chief AI scientist, Yao Shunyu, this month urged open-source and commercial developers to act responsibly as AI advances in biology — a reminder that the company frames openness as a managed franchise rather than a giveaway.

The competitive question is whether a 42.1 WorkIndex score at a twentieth of Hy4's cost pulls everyday office workloads away from Claude Haiku-class rivals. On Tencent's own blind-test numbers, satisfaction is a statistical tie and the price is not; if external benchmarks corroborate even part of the gap, the API economics of mid-tier assistants get interesting.

Source: mlq.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,070 articles

Related articles

« Previous articleNext article »