A VLM that answers with probabilities, not prose
Valen-Decision swaps text generation for structured decisions: feed it text, images or video plus your own candidates, get calibrated probabilities and bounding boxes back.
What it does
Valen-Decision is a family of multimodal models (0.8B to 27B) that act as a “System One” decision engine. Instead of generating an essay about an image, you hand it a task instruction and a set of candidates, and it returns probabilities over those candidates — plus bounding boxes via a dual-head design in the 1.5 release. It handles text, images and video, and the repo ships the full stack: model code, data processing, SFT and experimental RLCD training, and evaluation tooling.
The interesting bit
The core idea is replacing free-form generation with a typed decision interface — inspired by Typesafe’s Jev. That means no prompt-hacking a chatbot into pretending it’s a classifier; the model natively scores your options. The 1.5 release adds block-bidirectional attention and bbox output, so one forward pass can yield a choice, a confidence score, and a location.
Key highlights
- Five sizes from 0.8B to 27B, all Apache 2.0, in native, ONNX and GGUF formats
- Runs locally via a Valen-patched llama.cpp/Ollama that returns probabilities instead of text; also vLLM and SGLang
- Free hosted API with no key, plus Python, JS/TS and C++ SDKs
- Fine-tuning code included, with training on your own data supported
- Ships
VisualDecisionBenchfor image and video decision evaluation, and a Python SDKValen.llm()helper to benchmark against GPT/Claude/Gemini on the same questions
Caveats
- The RLCD training is explicitly labeled experimental
- GGUF support requires the custom patched llama.cpp runtime, not stock tooling
- The README doesn’t quantify the claimed state-of-the-art results inline; you have to dig into the evaluation figures and release posts
Verdict
Worth a look if you’re building pipelines that need fast, structured choices over images or video — moderation, routing, game agents, visual QA at scale. Skip it if you actually want a model that talks back.
Frequently asked
- What is Valen-Team/Valen?
- Valen-Decision swaps text generation for structured decisions: feed it text, images or video plus your own candidates, get calibrated probabilities and bounding boxes back.
- Is Valen open source?
- Yes — Valen-Team/Valen is open source, released under the Apache-2.0 license.
- What language is Valen written in?
- Valen-Team/Valen is primarily written in Python.
- How popular is Valen?
- Valen-Team/Valen has 933 stars on GitHub.
- Where can I find Valen?
- Valen-Team/Valen is on GitHub at https://github.com/Valen-Team/Valen.