---
title: "Tencent Gave Away a 770B Model. Selling Access to It Is Another Matter."
date: 2026-11-21
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-hunyuan-apache-tier
language: en
---

# Tencent Gave Away a 770B Model. Selling Access to It Is Another Matter.

> Tencent released full Hy4 weights on 20 November under a license derived from Apache 2.0, but added commercial thresholds that separate a permissive open model from a gated one.

A developer downloads a model, fine-tunes it on a laptop cluster and ships it inside a product. Another developer downloads the same weights and starts an API business. On paper both are "open source". In the license file, they are not the same customer at all.

That distinction is the real story of Tencent's Hy4 release. The flagship weights went public on 20 November 2026, completing the open release of the Hunyuan (混元) large model (大模型) series. But the text attached to those weights is not the plain Apache 2.0 licence that much of the early coverage described.

## What was released

Hy4 is a mixture-of-experts model with 770 billion total parameters, 49 billion activated per token and a context window above one million tokens. Its first release as a hosted service was on 4 November; the 20 November event was the publication of weights on GitHub, Hugging Face, ModelScope and GitCode. With it, the series — Hy3 preview, Hy3, Hy4 preview, Hy4 and Hy4 flash — became fully available.

The cheaper sibling, Hy4 flash, had opened earlier. Its Hugging Face card lists 220 billion total parameters, including about 64 billion "Engram" parameters, with roughly 12 billion active per token (11.3 billion for text and 0.43 billion for vision), a 1,048,576-token context and text-and-image input. Tencent launched it as an API on 16 November and uploaded weights days later, in instruction-tuned, FP8 and pretrained-base variants.

## Read the licence, not the headline

The Hy4 repository's licence file is titled "Tencent Hy License, Version 1.0". Its own preamble says it is derived from Apache 2.0 with modifications, mainly additional commercial terms for large model-as-a-service (MaaS) businesses in Section 10, and that it is not the Apache License and is not endorsed by the Apache Software Foundation.

Sections 1 through 9 are closely Apache 2.0, with clarifications that explicitly cover weights, parameters and configuration files, and define derivatives to include fine-tuned, quantized, distilled, merged and pruned versions. Section 4 adds an anti-circumvention rule: a downstream licence may not remove or limit Section 10, and Section 10 continues to apply to every recipient.

Section 10 is the gate. It applies to businesses that give third parties remote access to inference or fine-tuning — an API, in other words. Such a business needs a separate written licence from Tencent if either of two thresholds is crossed:

- aggregate worldwide gross revenue above US$50 million in any consecutive 12-month period, whether or not the revenue comes from the model; or

- a product using the weights with more than 100 million monthly active users.

Purely internal use is exempt, provided neither the model nor its outputs reach third parties. The definition also excludes end-user products where the model is confined to specific features, and mere relaying of requests to another provider.

Several technical outlets report that Hy4 flash carries the same licence. The earlier Hy4 preview, released in August, was under plain Apache 2.0 — which is likely why some Chinese coverage labelled the final model Apache 2.0 as well. If that label matters to your deployment, the licence file is the authority.

## Three tiers of "open"

A useful way to read the current Chinese open-model landscape is as three practical tiers. This is an analytical framing, not an official classification.

- **Permissive weights.** Apache 2.0 or MIT, with no scale-based gate. DeepSeek's models use MIT; the Hy4 preview was Apache 2.0.

- **Open weights with a scale gate.** The model can be downloaded and modified, but serving it commercially above a threshold requires a conversation with the vendor. Tencent's Hy License sits here, as does Alibaba's largest Qwen 3.8 model with a US$50 million rule; Moonshot's Kimi K3.1 Pro is reported to set the bar at US$20 million.

- **API only.** Capable models with no downloadable weights at all.

The middle tier is the interesting one. It lets a startup experiment without asking permission, while reserving the right to negotiate with the hyperscalers that would otherwise capture most of the value. It also makes "open weights" a weaker claim than "open source" in the strict sense.

## The numbers around the release

Tencent says Hy4 flash processed more than 50 trillion tokens on its first day and topped OpenRouter's daily leaderboard on 17 November. Those are company figures relayed by technology media and have not been independently verified. Listed API prices for Hy4 flash are ¥0.5 per million input tokens (≈ US$0.07 / HK$0.54), ¥2 per million output tokens (≈ US$0.28 / HK$2.17) and ¥0.02 per million cached input tokens (≈ US$0.003 / HK$0.02), all approximate.

The system card published alongside the weights is unusually candid. Tencent says training data for cyber security was restricted to defensive material such as proof-of-concept generation and vulnerability repair, with exploit and penetration content removed. It also reports regressions as well as improvements in its own tests, including higher scores on an exploit-writing benchmark and more misaligned behaviour under evaluation. Audits by Tencent's Xuanwu and Keen labs found no clear offensive-capability risk.

## Honest limitations

I could not open Tencent's licence file directly while researching this piece. The Section 10 terms above are reported by several independent technical publications that say they read the file, and are consistent with one another; readers making a deployment decision should open the file themselves. This is not legal advice.

Benchmark scores, first-day token counts and cost-efficiency ratios come from Tencent's own model card and launch material. No independent reproduction had appeared at the time of writing. The parameter breakdown differs between launch material (224 billion) and the Hugging Face card (220 billion); I have used the card. Coverage disagrees on whether the final Hy4 is Apache 2.0 or Tencent Hy License, and the disagreement itself is the reason to check the source file.

The three-tier framing is mine, for orientation, not a scheme published by any company or standards body.

## What readers can do now

- **If you plan to serve a model to others, read Section 10 before you write any code.** The thresholds count total company revenue, not revenue from the model.

- **If you only use a model internally, the extra terms generally do not apply.** Document that the outputs stay inside your organisation.

- **If you are comparing models, treat "open weights" and "permissive licence" as two separate questions.** A model can be freely downloadable and still require a commercial conversation.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-hunyuan-apache-tier
Free to quote with attribution and a link to the original.
