// AI NEWS · MODEL RELEASE

Tencent Open-Sources Hy4 Preview, a 770B-Parameter Agentic Model Built for Coding and Research

On August 28, 2026, Tencent Hunyuan released and open-sourced Hy4 preview — a mixture-of-experts model with 770B total parameters, 49B active, and a context window over 1 million tokens, aimed at software engineering, office work, and scientific research.

2026-08-29 · by Moe Ameen

What happened

On August 28, 2026, Tencent's Hunyuan team released and open-sourced Hy4 preview, a next-generation large language model. It is a mixture-of-experts (MoE) design with 770 billion total parameters and roughly 49 billion active per step, paired with a context window Tencent describes as exceeding 1 million tokens. The stated target is productivity work — software engineering (debugging and long-context code tasks), office and financial analysis, game development, and scientific research including AI R&D and physics simulation — rather than chat or creative content.

Tencent published the model weights openly on Hugging Face, ModelScope, GitCode, and CNB, and also shipped a lower-precision Hy4 preview-FP8 variant alongside the full-weight release. The model is available inside Tencent products including WorkBuddy, CodeBuddy, Yuanbao, and ima, and via API through Tencent Cloud TokenHub and OpenRouter. Tencent listed API pricing at about $0.834 per million input tokens and $2.501 per million output tokens, with a limited free-access window through WorkBuddy and CodeBuddy at launch. Confirm current pricing and access terms on Tencent's own pages before you budget around them.

Tencent reported gains across a set of engineering benchmarks and, in an internal blind evaluation, an average score of 2.99 out of 4.00 across 203 engineering tasks judged by 163 experts — slightly ahead of the GLM 5.3 and Kimi K3 numbers it cited. Because that blind evaluation is Tencent's own, treat it as a vendor claim, not an independent result. The most unusual detail in the announcement is agentic: Tencent says Hy4 participated in optimizing its own training methods, data strategies, and evaluation frameworks, and autonomously tuned its inference system for a reported 31.8% end-to-end throughput improvement.

Why it matters for creators

  • This is a coding-and-reasoning model, not a content studio. Hy4 is text-in, text-out — no images, video, audio, or publishing. For creators it is a drafting and research brain, not a tool that makes finished posts.
  • Open weights at frontier scale keep pushing the cost of raw text toward zero. A 770B MoE you can self-host or hit cheaply via OpenRouter means high-quality drafting and long-context analysis are no longer the scarce part of a content operation.
  • The 1M-token context is the creator-relevant strength. It can ingest an entire webinar transcript, a research corpus, or a quarter of documents in one pass and return a structured brief — useful upstream, before any content gets made.
  • Its tuning is engineering-first, so don't assume it writes great captions. Hy4 is optimized for code, agentic tasks, and analysis; its English marketing-copy and short-form voice are unproven, and a raw model draft still reads like a raw model draft.
  • As open text gets commoditized, the differentiator moves downstream — brand consistency, format breadth, and reliable multi-platform distribution — not which model wrote the first draft.

How to act on this with Kompozy

There are two ways to act on this today. The fast one is to ride the story: paste "Tencent just open-sourced Hy4 preview, a 770B agentic coding model" into [Kompozy](/)'s Quick Ingest and fan that single beat into a Text Post, an X thread, a LinkedIn [Carousel](/glossary/hyperframes), a captioned [Persona Short](/glossary/persona-shorts) explaining what open weights mean for builders, and a blog explainer — then schedule and publish the set across the eight social platforms plus blog and email. Kompozy is a generation-and-publishing engine, so one news item becomes a full cross-platform package the same afternoon.

The more durable move uses Hy4 for what it is actually good at. Point it at a dense source — a long technical transcript, a set of release notes, an earnings call — and let its million-token context return a clean brief or a slate of scripts. Then drop the strongest draft into Kompozy as a source, and Kompozy does the part a text model can't: it generates roughly 25–35 finished assets across 18 formats, rewrites every one under a [Persona Brief](/glossary/persona-brief) so it sounds like you instead of like a raw model, puts a face on the dry material with a HeyGen avatar, and ships the batch through [Autopilot](/glossary/autopilot) behind a per-post review. Hy4 reasons and drafts at the front; Kompozy builds the identity and delivers it to an audience.

Quick takeaways

  • Tencent Hunyuan open-sourced Hy4 preview on August 28, 2026 — a 770B-parameter MoE model with ~49B active and a context window over 1 million tokens.
  • Open weights (plus an FP8 variant) are on Hugging Face, ModelScope, GitCode, and CNB; it is also in WorkBuddy, CodeBuddy, Yuanbao, and ima, with API access via Tencent Cloud TokenHub and OpenRouter at ~$0.834/M input and $2.501/M output tokens.
  • It targets coding, office and financial analysis, game development, and scientific research — a productivity model, not an image/video/audio generator.
  • Tencent reports strong engineering benchmarks and a 2.99/4.00 internal blind-eval average; treat vendor benchmarks as claims until independently verified.
  • For creators, Hy4 is a drafting and long-context research engine — pair it with Kompozy to turn its text into on-brand, multi-format, published content.

Frequently asked questions

What is Tencent Hy4 preview?

Hy4 preview is a large language model released and open-sourced by Tencent Hunyuan on August 28, 2026. It uses a mixture-of-experts architecture with 770 billion total parameters (about 49 billion active per step) and a context window exceeding 1 million tokens, and it is built for software engineering, office and financial analysis, game development, and scientific research.

Is Hy4 preview really open source, and where can I get it?

Tencent published the Hy4 preview model weights openly — along with a lower-precision Hy4 preview-FP8 variant — on Hugging Face, ModelScope, GitCode, and CNB. It is also usable inside Tencent products like WorkBuddy, CodeBuddy, Yuanbao, and ima, and via API through Tencent Cloud TokenHub and OpenRouter. Confirm the exact license terms on Tencent's release pages.

Can Hy4 preview generate images or video for content?

No. Hy4 is a text-in, text-out model focused on coding, reasoning, and analysis — it does not produce images, video, or audio, and it does not publish. To turn its drafts and briefs into finished visual content and posts, pair it with a generation-and-publishing engine like Kompozy, which makes persona/avatar video, carousels, images, blogs, and newsletters and schedules them across platforms.

How much does the Hy4 preview API cost?

Tencent listed API pricing at roughly $0.834 per million input tokens and $2.501 per million output tokens, with a limited free-access period through WorkBuddy and CodeBuddy at launch. Prices and access windows can change, so verify current figures on Tencent Cloud before relying on them.

Related news

← All AI news · Get started →