Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

dfip

Research preview: Python client and pinned adapters for running System One decision models on your own Modal GPUs.

LibraryresearchMIT
View on GitHub ↗

Overview

DFIP (decision-focused inference) is a research preview that provides a small Python client, pinned model adapters and reproducible experiments for System One decision models, with weights and GPU execution kept in your own Modal account. The client runs locally, setup downloads only the selected pinned checkpoint into a Modal volume and deploys authenticated methods, and no public HTTP endpoint is created. The quick start uses an L40S, and containers scale to zero after a minute of inactivity.

The client takes Jev-style requests (state plus typed questions such as a team choice) and returns typed answers. The default deployments use released native inference paths; CUDA graph replay, Triton candidate projection and shared batching are separate research paths that the README says show gains on some workloads but do not establish a general acceleration method or production readiness. Model licences, such as Liquid's commercial-use conditions, apply independently.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich team should handle this request?—

How it works

  1. dfip setup deploys a pinned checkpoint to Modal
  2. Client sends state and typed questions
  3. Model runs on your Modal GPU
  4. Typed answers return to the local client

Strengths

  • Weights and execution stay in your account
  • Pinned, reproducible checkpoints
  • Scale-to-zero GPU usage

Caveats

  • Research preview, not production-ready
  • Requires a Modal account with GPU access and incurs usage
  • Model licences vary

Getting started

pip install -e ., modal setup, then dfip setup --model liquid and dfip predict --model liquid --input examples/request.json.

At a glance

Type
Library
Status
research
Language
Python
Runtime
Modal GPU functions
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

Frequently asked questions

Where do models run?

On GPUs in your own Modal account.

Is it production-ready?

No, it is a research preview.

Which models are supported?

Pinned adapters such as a Liquid decision model; see the repository for the list.

Related projects

  • go-system-one

    Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.

    ProjectbetaGo
  • hearim

    Go gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.

    ServicebetaGo
  • llm2jev

    Turns any chat model on SGLang, vLLM, transformers or MLX into a Jev-compatible service with one prefill and one label token per question.

    ProjectbetaPython
  • verdict

    Self-hosted Jev-compatible System One server that sits in front of llama-server and reads option probabilities from any GGUF model.

    ServicebetaPython
  • jevinf

    Inference engine for Jev-style decision models using segmented forwards with prefix reuse, serving the Jev wire contract.

    ProjectresearchPython
  • simple-jev★ 602

    Featherless AI server that turns compatible open Hugging Face models into a /v1/systemone classifier by reading next-token logits.

    ServicebetaPython