Valen brings fast-twitch visual reflexes to AI agents

Valen is an open-source multimodal decision model designed to act as a fast reflex layer for AI agents. Instead of generating wordy text explanations through slow autoregressive decoding, it inspects text, images, or video and directly scores predefined action choices with probability distributions.

Most modern agent pipelines make the mistake of routing every visual perception task through a massive vision-language model. Valen is built explicitly for System One intuitive decisions, handling the repetitive, split-second judgments where standard models are simply too slow and expensive.

What it does and how it works

Valen takes raw perceptual state alongside task instructions and returns structured probabilities across candidate answers.

It handles three native decision patterns. Predicates return a zero-to-one truth probability for a given condition. Choice queries output normalized probabilities across arbitrary options. Score queries calculate a weighted score based on ordered semantic levels alongside a confidence metric.

That makes it ideal for low-latency agent loops. GUI agents can check if a modal pop-up is blocking the screen or pick the right button to click. Game agents can navigate mazes, Sokoban puzzles, or lane-racing tracks in real time. Mobile robots can instantly decide whether to stop, skirt around, or push through an obstacle.

In customer ticket routing tests, adding a fallback option allowed Valen to catch out-of-scope requests and pass them to humans, cutting false automated approvals from eight down to one across two hundred test cases.

Specs, speed, and running it

Under the hood, Valen relies on Qwen3.5 backbones paired with a Mixer architecture and shared-state inference, trained on millions of supervised fine-tuning examples.

It comes in three sizes: 0.8B, 2B, and 4B parameters. Because it skips autoregressive text generation entirely, single-step decisions clock in at roughly 122 to 128 milliseconds, while four parallel game rollouts resolve in under 1.24 seconds.

The team open-sourced full weights on Hugging Face, modular ONNX exports with Python, JavaScript, and C++ SDKs, and custom GGUFs compatible with patched builds of llama.cpp and Ollama. For production serving, Docker builds are provided for both SGLang and vLLM.

You do not need local GPUs just to try it out. The project hosts a free public API endpoint and a web demo that requires no API key. They even shipped an agent skill for Claude Code and Cursor, plus a benchmarking suite called JevalKit to pit Valen directly against frontier models like GPT, Claude, and Gemini.

Why it matters

If you are building computer-use agents, robotics software, or game automation, Valen is worth testing right away. Calling an omni-scale frontier model just to confirm whether a button is active introduces unnecessary latency and billing overhead. Valen slots neatly into the loop as a lightweight reflex engine.

We think divorcing decision-making from token generation is the smartest architectural trend in agent design right now. The catch is that Valen requires you to define candidate spaces in advance; it will not invent creative solutions out of thin air. But for high-frequency, bounded decisions, letting a small model handle split-second instincts is exactly how agent stacks become practical.