Hunyuan Hy4 Preview Workbuddy Kaihe Aibox

The Short Version

📖 Glossary

AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

On August 28, Tencent Hunyuan released and open-sourced its new flagship model, Hy4 preview: 770B total parameters, 49B active, and a context window exceeding 1M tokens. It landed first on WorkBuddy (both China and international builds) with a two-week free trial. For KAIHE AIBOX users, the signal is simple: the stronger, cheaper, and more open the models get, the more ammo your local agent computer can call on.

The Misconception: This Isn't Just "Bigger Parameters"

Most people see "770B" and scroll past, assuming it's another spec-shouting release. But Hy4 preview's interesting part isn't the raw number — it's the positioning and the structure.

It uses a MoE (Mixture-of-Experts) architecture: 770B is the "capacity reserve," but only 49B activates per response. In plain terms, the model packs far more knowledge into its belly, yet when actually working it only wakes the small slice of parameters relevant to the current task. This is the same philosophy as KAIHE AIBOX — don't brute-force with hardware; call the right thing at the right time. The cost stays bounded while the capability ceiling rises.

Context jumped from Hy3's 256K to 1M — roughly four times more material in one shot. A full manual, an entire code repository, or a stack of papers can go in at once without slicing. For long-document analysis, cross-file refactoring, and multi-turn complex tasks, that's a real difference, not a benchmark game.

Change 1: Built to Work, Not to Answer

Hunyuan positions Hy4 preview as "built for productivity." Its training data wasn't scraped loosely from the web — it was co-built with Tencent's own software-engineering, gaming, finance, and security experts, then tuned alongside products that actually do work for people, like WorkBuddy and CodeBuddy (what they call Co-Design).

Concrete scenarios: - Software engineering: stronger long-horizon planning, debugging, and verification, plus better front-end aesthetics. Terminal-Bench 2.1 rose from 70.8 (Hy3) to 85.4. - Office analysis: deeper understanding of complex work environments and financial analysis, optimized for data analysis and cross-file collaboration — from reading material all the way to delivering documents, spreadsheets, and presentations. - Game development: one sentence of requirements generates a playable prototype and drives a game engine; developers refine through multi-turn interaction. - Scientific research: clear progress in AI R&D, molecular dynamics, condensed-matter physics, and fundamental mathematics.

The sharpest move: Hy4 preview participated in its own development for the first time — proposing approaches, running experiments, reading results, feeding code and logs back into the next training round, forming an early recursive self-improvement loop. It also optimized its own inference system, lifting end-to-end throughput by 31.8% over baseline.

Change 2: WorkBuddy First Integration, Two-Week Free Trial

Hy4 preview launched simultaneously on WorkBuddy (China and international) and alongside Yuanbao, ima, and CodeBuddy. It shipped with a "limited free" tag on day one; the free window runs to 2026-09-10 23:59. The previous Hy3 free period was extended to September 30 as a fallback for existing users.

Demand was so high that queues appeared on day one, and Tencent urgently expanded its inference cluster three days later — though peak hours may still wait. One detail worth noting: Hy4 is a text-only language model and is not multimodal. Multimodal tasks in WorkBuddy auto-route to other models and consume credits, not Hy4's free quota.

Pricing stays accessible: input (cache hit) ¥0.3, input ¥6, output ¥18 per million tokens. Against the recent wave of price hikes elsewhere, that's friendly for heavy users.

The KAIHE AIBOX Angle: Stronger Models Make the Local Base More Valuable

Let's draw a hard line here: WorkBuddy is Tencent's cloud product — KAIHE AIBOX doesn't run the WorkBuddy cloud office suite. But Hy4 is a cloud model API, and it's precisely the kind of thing KAIHE AIBOX connects to: an agent base pulling in cloud models is exactly what it's built for. Don't confuse "cloud office product" with "local agent computer."

But the two are complementary. KAIHE AIBOX is a locally-running agent computer — plug in and it works, 7×24 low-power standby, data stays in your home. Its real value isn't binding to one model; it's freely scheduling external model APIs. Hy4 comes in via Tencent Cloud TokenHub or OpenRouter, while the local agent orchestration handles task decomposition and keeps your workflows stored locally — the model is just one tool being called.

In other words, the stronger, cheaper, and more open cloud models get, the bigger the local base's upside. You don't swap hardware because "some model updated again." The Hermes agent orchestration on KAIHE AIBOX lets you switch models like swapping a battery. Use Hy4 today, switch to another flagship tomorrow — your workflows stay in your hands.

How to Actually Use This Update

If you already use WorkBuddy for complex reasoning, long-horizon development, or office analysis, these two weeks (until Sept 10) are a zero-cost window to test the 770B flagship on real projects — not just small writing. For light daily Q&A, keep Hy3 free as the fallback; when Hy4 queues or its trial ends, Hy3 stays free through September.

For KAIHE AIBOX users, my advice is sharper: watch the "local orchestration + free multi-model switching" capability. Models will keep changing generations, but the baseline — workflows沉淀 locally, data staying at home — won't shift because of any single cloud release. That's the most fundamental difference between KAIHE AIBOX and pure-cloud solutions.