Tencent Opens Hy4 Flash Weights, With a Catch for Big Sellers

Human Verified

Share to: , and

A glowing neural network cube floating above an open laptop on a desk at night

Tencent published the weights for Hy4 flash, its most cost-effective model to date, as open-weight releases this week, two days after the model reached its own API. The release covers three checkpoints: the instruction-tuned model, an FP8 build and a pre-trained base.

What Hy4 Flash Is

Hy4 flash is a mixture-of-experts model with 220 billion total parameters, including 64 billion in Engram n-gram embedding tables, and just 12 billion active parameters per token. It takes text and image inputs, generates text, and supports a context window of up to one million tokens.

Tencent built the model for cheap everyday work: deep search, document processing, slide creation and simple computer use. Its own WorkIndex benchmark scores it at 42.1, slightly above the larger Hy4 preview, at roughly a thirteenth of the cost per task.

The Licence Is the Story

The weights ship under the Tencent Hy License 1.0, an Apache 2.0 derivative with extra commercial terms in Section 10. Anyone running a model-as-a-service business above 50 million dollars in annual revenue must obtain a separate written licence before using the weights.

The same rule applies to any product or service with more than 100 million monthly active users, while purely internal use is exempt. The download is free, in other words, but deployment at scale is not unconditional — and Tencent named a licensing contact in the licence itself.

How It Performs

In a blind evaluation with 85 of Tencent’s own WorkBuddy users across 259 tasks, Hy4 flash scored 2.79 out of 4 on satisfaction, close to Claude Haiku 5.5 at 2.81. It recorded the lowest cost per task and the shortest median completion time, at 6.7 minutes.

The model card is candid about limits: heavy synthetic instruction data in mid-training can leave the base model struggling to follow expected answer formats in few-shot tests. Transformers support is still landing through a pull request, not a merged release.

For serving, Tencent recommends vLLM or SGLang with its FP8 build, claiming decoding past 150 tokens per second on an eight-way tensor-parallel node. A built-in drafter, DFly2, reuses the target model’s key-value cache to keep long-context decoding cheap at scale.

Our Read

Hy4 flash is Tencent’s bet that cheap beats clever: a model distilled for everyday productivity, priced to undercut rivals and now free to run on your own hardware. The blind-test draw with Claude Haiku 5.5 is the headline, but the cost figures are the argument.

The licence is the catch that matters. A 50-million-dollar revenue threshold and a 100-million-user line keep the weights open for researchers and small operators while reserving the big commercial upside for Tencent. Open weights, in this release, does not mean open business.

Why It Matters

Open-weight releases from China’s big labs are now arriving within days of the hosted launch, and Hy4 flash is the latest proof. For builders, a 220-billion-parameter model at a thirteenth of the cost of its predecessor changes what can be self-hosted — if the licence terms allow it.

Background

Tencent introduced Hy4 on November 4 as a model that can see and work the computer, and that launch focused on agentic computer use. Hy4 flash distils that larger model into a cheaper, faster package aimed at everyday productivity tasks, following Qwen4-Max in the open-weights race.

Source: Tencent’s model card and its licence

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.