Fast Decisions. Continuous Evolution.
Fast decision models · Recursive self-improvement (RSI) · Agents
English · 简体中文
Website · Playground · Documentation · News · Hugging Face · Leaderboard
Valen is an open-source organization bringing together fast decision models, recursive self-improvement (RSI), and agents. We develop and bring together models, tools, research resources, and applications to help agents make decisions quickly, act in their environments, and improve through experience and feedback.
Our goal is to connect efficient decision-making with agents that can learn and adapt over time. We welcome researchers and developers working across these areas to build, evaluate, and share their work here.
- Fast decision models — low-latency decisions from text, images, and video, with structured outputs that agents can use to select actions, route tasks, and assess states.
- Recursive self-improvement (RSI) — methods for learning from experience, reusing memory and skills, improving tools and agent workflows, and updating models. We care about reproducible evidence that these changes lead to better performance over time.
- Agents — agent frameworks, tools, and applications that connect perception, planning, decision-making, and execution in interactive environments.
These directions inform one another: decision models help agents choose actions; agent runs produce experience and feedback; RSI methods explore how to turn that feedback into improvements. Evaluation helps establish whether those improvements hold up.
| Project | What it does | Explore |
|---|---|---|
| Valen-Decision | A multimodal System One decision model: text, images, and video in; structured decision probabilities out. Valen-Decision-1.5 comes in five sizes from 0.8B to 27B and answers bbox questions. Includes training, fine-tuning, inference, and evaluation. | Try it · Playground · Docs & API |
| JevalKit | An evaluation toolkit for Jev-like decision models, with unified model interfaces, prepared datasets, benchmark scoring, and a public leaderboard. | Docs · Leaderboard |
| Awesome-RSI | A curated collection of papers, code, benchmarks, datasets, and models for recursive self-improvement, organized around experience accumulation, system modification, and model parameter updates. | Papers · Projects |
Our current public projects cover decision models, evaluation, and RSI research resources. We plan to expand the organization with more agent and RSI projects, sharing implementations, experiments, and reusable tools as they become available.
New: Valen-Decision-1.5 is released in five sizes, as native weights, ONNX and GGUF. Get it from the Valen-Decision-1.5 collection.
Browse all releases on Hugging Face, or start with the documentation to use Valen-Decision and run evaluations with JevalKit.
Valen is named after evolutionary biologist Leigh Van Valen, who proposed the Red Queen hypothesis: organisms must continually adapt as the species around them evolve. Its literary inspiration comes from the Red Queen's race in Through the Looking-Glass.
For us, the name connects speed with continuous evolution: making decisions quickly while continuing to learn and adapt.
Our Chinese name, 万澜 (Wàn Lán), evokes countless waves in motion. The wave-shaped V in our logo carries that idea forward.
We welcome contributions and collaboration across decision models, RSI, and agents:
- Decision models and evaluation — improve Valen-Decision, add models or benchmarks to JevalKit, or share data and reproducible results.
- RSI research and methods — contribute approaches to learning from experience, evolving agent systems, or updating models; suggest papers and implementations for Awesome-RSI.
- Agents and applications — contribute agent frameworks, tools, integrations, and applications, or propose new agent and RSI projects for the organization.
Open an issue or submit a pull request in the relevant repository. Documentation, examples, and translations are welcome too.