📖 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.
Summary: On November 4, Tencent Hunyuan released Hy4 with 770B parameters, first vision, first computer use. The dividing line is not a leaderboard spot but the reach of an agent: from systems with APIs to everything visible on a screen. Once capability gets there, the bottleneck moves to the person, which is where KAIHE AIBOX（铠盒AIBOX）sits.
What actually shipped
Hy4 is the official release, dated November 4. 770B parameters, mixture-of-experts, 1M context. Against the August preview it scales pretraining and post-training and adds two firsts.
Vision understanding arrives first: detail in an image, a complex chart, visual reasoning folded into math problems.
Computer use arrives for the first time: a command line and a graphical interface both. When a job spans several applications it clicks through the screen and types into the terminal itself. Tencent built a sandbox with several hundred applications and let the model operate repeatedly, checking its own results. Blender and CAD got dedicated training.
Before, you asked and it answered. Now you hand it a goal and it starts working.
The word that changed is "interface"
Agent talk assumed an API for two years. The model calls a tool, the tool returns a result, the loop closes neatly.
Real workplaces do not look like that. Plenty of systems that still run businesses have no API: old ERP installs, internal admin panels, professional software on a desktop. Their interface is a screen, a mouse, a keyboard.
Computer use stretches what an agent can touch from systems with an API to anything visible on screen. Not one capability added. A change of scale.
Three numbers, one underlying ability
The published evaluations look unrelated. All three measure running long, seeing the state, verifying itself.
Game development: integration with UE, Unity and Godot, built with professional developers at Tencent IEG. GameDev up 19.1% and GameLogic up 9.8% against the preview.
Financial analysis: 213 experts blind-testing 579 real tasks. Hy4 scored 3.04, first among open models. DeepSeek-V4.1-Flash-1017 sat at 2.97; Kimi K3.1 Pro and GLM-5.3 both at 2.68.
Error rates: hallucination fell from 12.95% in the preview to 7.2%; errors from guessing fell from 23.1% to 5.6%.
The shared trait is not the score. It is a dozen linked steps, each checked. A model that only answers questions has no reason to push hallucination to 7.2%. A model that clicks a mouse and must confirm the click landed does.
Capability moved. The bottleneck moved with it
Task length follows from a model that keeps working. Longer tasks put the person in an awkward spot.
| Task | Typical runtime | Someone watching? |
|---|---|---|
| Look up a fact | seconds | no |
| Draft a document | fifteen minutes | a glance |
| Install an environment, edit files across three apps | one to two hours and up | yes |
| Cross-check records for several days | measured in days | yes, machine left on |
The last row is where things stand. Model capability stretches across days. Your computer is only on while you sit at it.
The limits the vendor drew itself
Hunyuan's own framing: Hy4 sits among the strongest open models, a real gap remains to the top closed models. Biosecurity open-ended experimental design scored 30.4 of 100, and no meaningful lowering of the bar for harmful biochemical work was found. Official weights are not out yet; Tencent says soon. Entry points live: TokenHub, WorkBuddy, CodeBuddy, Yuanbao, ima; the API also landed on OpenRouter, Hermes, OpenCode.
Where this lands for an ordinary user
Model rankings are not your problem; Tencent pays for the research. Your problem is narrower: getting the work done without sitting there while it happens.
KAIHE AIBOX occupies that slot. A dedicated agent computer: agent system built in, running around the clock on its own, without tying up your laptop. You set tasks in the day, it executes at night. Cloud models cost less off-peak, and the agent schedules work toward those windows, so the difference lands on your bill. Hardware and agent system bought once, no monthly software subscription.
The two rows reading "yes" collapse into one sentence: go to sleep, it is handled.
Once AI grows a mouse, the scarce thing stops being intelligence. It is whether some machine will sit through those hours on your behalf.
Related Reading: - Does My AI Work Without My Computer? - Too Many Errands to Remember? - Local AI Assistant vs Chat AI
KAIHE AIBOX #Tencent Hunyuan #Hy4 #AI frontier #agent computer #local AI
For more information, search [KAIHE AIBOX] or contact: [email protected]
KAIHE AIBOX · 7x24 Personal AI Assistant | AI前沿