GeekPark: From Jev to StartLux, How AI Decision Models Fill the 'Intuition Layer' for Local Deployment
A GeekPark article analyzes AI decision models ("System One") represented by TypeSafe's Jev, which decouple high-frequency judgment from deep reasoning. It highlights that while the category has low barriers and rapid iteration, Chinese teams like Shanghai AI Lab and StartLux are filling infrastructure gaps for data-sensitive scenarios through open-source and localization strategies, enabling complex Agents to run on personal computers.
In practical AI Agent applications, what occurs densely is often not long-chain deep reasoning, but a multitude of trivial judgments: Which team should handle this ticket? Can this command be released? Is the task complete? Previously, these judgments were either handled "incidentally" by expensive large models or covered by rigid rules—the former slow and costly, the latter unintelligent. TypeSafe AI's Jev model proposes a new architectural approach: stripping "judgment" from "generation" to create a cheap, high-frequency infrastructure. This type of model, referred to as "System One," only answers questions with preset options, directly returning structured results with probabilities at a very low price point.
Jev's success proved the market value of this path, but its closed-source hosted nature limited its application in scenarios requiring data sensitivity or local deployment. This left an entry point for followers. Recently, global giants including OpenAI, Cloudflare, and Amazon have followed suit with similar APIs or models. Meanwhile, Chinese teams rapidly adopted a differentiated strategy: open-sourcing and localization. Shanghai AI Lab open-sourced the multimodal decision model Intern-Decision, while StartLux (Yuan Dian Xing Hui), a Shanghai company founded just five months ago, released StartLux-Decision with support for quantization files, offering specifications ranging from 0.8B to 27B, aiming to keep model capabilities and data on users' own devices.
For companies betting on "Local AI," decision models are not a hot new category to chase, but a critical component for whether local systems can run effectively. Cloud computing power can be stacked, but personal computers have hard limits on VRAM, memory, and power consumption. If every small judgment requires a 27B parameter large model, local Agents simply cannot run efficiently. Data published by StartLux shows that its 4B version, after Q4 quantization, requires only about 2.71GB of VRAM, maintaining a 98.3% consistency rate with original weights. This layered architecture—large models handling general capabilities, small models handling high-frequency judgments—makes building a complete personal AI system on resource-constrained devices possible.