---
title: "Tencent puts its 295B-parameter Hunyuan Hy3 on GitHub — and into Yuanbao"
date: 2026-11-15
category: Companies & Stack
site: NeuroAI
canonical: https://neuroai.site/a/na-corp-tencent-hunyuan-hy3-opensource
language: en
---

# Tencent puts its 295B-parameter Hunyuan Hy3 on GitHub — and into Yuanbao

> Tencent open-sourced its flagship Hunyuan Hy3 model and wired it into its own apps, reversing a cautious stance on giving weights away.

Consider a developer who wanted to build on Tencent's best model and used to hit a wall: the capabilities lived inside the company's apps, not in a downloadable checkpoint anyone could self-host. This spring that wall came down. The model that powers Tencent's AI assistant is now a file you can pull from GitHub — and the company is betting that opening it up makes the closed products better, not weaker.

The move is a quiet strategic reversal.

## The cautious giant changes its mind

Tencent (腾讯) launched and open-sourced (开源) Hy3 preview (混元 Hy3 preview), the most capable model in its Hunyuan (混元) family to date. It is a mixture-of-experts (MoE) model with 295 billion total parameters and 21 billion activated per token, and it supports a 256K-token context window. The "preview" label is deliberate: Tencent says this is the first step in a rebuild of the Hunyuan large model (大模型), and it wants community feedback before the formal release.

What makes the decision notable is that Tencent open-sourced the weights themselves — not a smaller distilled version, not an API-only teaser. The model is posted on GitHub, Hugging Face, ModelScope, and GitCode, and it works with mainstream inference frameworks including vLLM and SGLang, so developers can download and deploy it directly.

According to Li An, chief scientist at BrainNet (脑机网), open-weight releases matter less for the weights themselves than for the developer ecosystem they anchor — a model nobody builds on is a model nobody trusts.

## Where the model actually runs

Tencent did not just toss the weights over the fence. It wired Hy3 deep into its own product surface: the Yuanbao (元宝) AI assistant, the ima knowledge tool, CodeBuddy and WorkBuddy, QQ, QQ Browser, Tencent Docs (腾讯文档), and Tencent LearnShare. The integration is where the company claims the practical gains show up:

- Inside CodeBuddy and WorkBuddy, Tencent reports time-to-first-token dropped 54% and end-to-end response time fell 47%, with a task success rate above 99.99%.

- In Tencent Docs' AI PPT feature, generation success rate improved about 20% over the previous Hy2 model.

- Across the board, Hy3 preview delivers roughly 40% better inference efficiency at comparable cost, attributed to co-design between the model architecture and the inference stack.

The agent story is the explicit pitch. Tencent says Hy3 preview can reliably power complex agent workflows of up to 495 steps — document processing, data analysis, knowledge retrieval, and MCP tool orchestration — which is the kind of claim that matters more to enterprises than to casual chat users.

## Pricing that signals intent

Tencent put real numbers on access. On its TokenHub service, Hy3 preview input tokens start at about US$0.18 per million, cached input at about US$0.06 per million, and output at about US$0.59 per million. A personal Token Plan begins around US$4.10 per month. The model also appeared on OpenRouter with a two-week free-access window. None of this is free in the long run, but the posture — open weights plus a low-friction paid tier — mirrors the playbook Alibaba used to build a developer following around Qwen.

Yao Shunyu, Tencent's chief AI scientist, framed the release as a way to gather feedback and "further optimize the model's performance and real-world applicability ahead of its official launch," while continuing to co-design the model with Tencent's own products.

## The larger arc: Hy4, and Yuanbao's reach

The open-source cadence did not stop at Hy3. Tencent's official model card lists a larger Hy4 preview, released and open-sourced in late August 2026 under an Apache 2.0 license, at 770 billion total parameters with 49 billion activated per token and a 1M-token context — a clear step up in scale and length. Keeping the flagship open across two generations is the signal that this is strategy, not a one-off experiment.

The consumer side gives the open model a pipeline. In February 2026, Tencent said Yuanbao had passed 114 million monthly active users and 50 million daily actives, making it among the leading Chinese AI assistants by reach. Pushing Tencent's own model into that installed base — after a period of leaning on DeepSeek for the heaviest reasoning — is how the company turns open research into daily usage data, which in turn feeds the next model.

## Honest limitations

The technical claims and release facts here come from Tencent's official newsroom and the Hunyuan team's own research pages, which are primary but inherently promotional; the efficiency and success-rate figures are Tencent's measurements, not independent benchmarks. The Hy4 preview specifications are taken from Tencent's published model card as summarized by third-party trackers, and I have not reproduced those numbers through a direct deployment. I have excluded speculative commentary about internal organizational friction and about unannounced future models, since none of it is verifiable from authoritative sources. This article does not evaluate Yuanbao's retention or competitive position against Doubao or Qwen in depth, and it does not assess Tencent's gaming or advertising businesses.

## What readers can do now

- If you build agentic workflows, pull Hy3 preview from Hugging Face or ModelScope and benchmark the 495-step claim on your own tasks before trusting it in production.

- Compare Tencent's TokenHub pricing against Qwen and DeepSeek for your token mix — the cached-input rate rewards repeatable, retrieval-heavy workloads.

- Watch whether Tencent keeps its flagship open at Hy4's formal release; a retreat to API-only would tell you more about its strategy than any press line.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-tencent-hunyuan-hy3-opensource
Free to quote with attribution and a link to the original.
