this-that-model
Open 1.9B typed-decision model that returns an option index and calibrated probability in one forward pass, with a paper and API.
Overview
this-that-model is an open-weight decision model from FLock with an accompanying arXiv paper. It takes a state string and a question with named options and returns an index and a probability per option in a single forward pass, with no decoding loop, parser or retries. The architecture is 1.88B parameters, mostly DeltaNet linear attention layers.
It targets developers who hit a branch they cannot express in code, such as whether a refund is within policy or whether a shell command is safe to run unattended. The README reports 31.4 ms per decision on a consumer card and offers free use through the FLock API, plus a /v1/systemone endpoint. Three checkpoints (1.0, 1.1, 1.2) are published, with tables comparing composed-decision accuracy and calibration. The README is candid that out-of-distribution questions produce low-confidence or wrong answers and should be thresholded and routed elsewhere.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Is this shell command safe to run unattended on a production host? | yes, it only reads state · no, it modifies or deletes data · no, it contacts the network |
| Yes / no | Should this request be rate-limited? | yes · no |
How it works
- Load the model with TypedDecider.from_pretrained or call the FLock API.
- Pass state text and a Question with named options.
- Read the argmax index and per-option probabilities.
- If confidence is below a threshold (e.g. 0.8), escalate to another handler.
Strengths
- Single forward pass with no generated tokens.
- The README reports 31.4 ms per decision on local hardware.
- Calibrated probabilities make thresholds meaningful, with published calibration figures.
- Weights, benchmarks and scripts are public for reproduction.
Caveats
- Out-of-distribution questions can return confident-looking wrong answers; the README shows an example.
- Later checkpoints trade some calibration for accuracy per the published table.
- Benchmark figures for 1.1 and 1.2 come from model cards, not this repo's scripts.
Getting started
from thisthat import TypedDecider, Question; decider = TypedDecider.from_pretrained("flock-io/this-that-model-1.2")At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- MLX / PyTorch, hosted API
- Has code
- yes
Use cases and integrations
How curated lists describe it
A 1.9B typed-decision model with an arXiv paper behind it: one forward pass, no decoding loop, and a `/v1/systemone` endpoint.
MrJev/awesome-jev
Frequently asked questions
How big is the model?
About 1.88 billion parameters, with 18 of 24 layers using DeltaNet linear attention.
Can I use it without a GPU?
The README offers free access through the FLock API as an alternative to running it locally.
How is it different from asking an LLM?
It scores the given options in one pass and returns an index and probabilities instead of generating text that must be parsed.
Related projects
Exact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPythonLocal reproduction of Jev-style inference that reads option-token logits from a small LLM in one pass, with an honest failure log.
ProjectresearchPythonOpen decision models that derive calibrated confidence from recurrent looped-layer dynamics instead of RL, with early exit and a Jev API.
ProjectresearchPythonLaptop-scale reimplementation of RL for Calibrated Decisions: parallel constrained decoding trained with GRPO and a Brier reward.
ProjectresearchPython- AnyJev★ 1,255
Nokia's training-free library that turns open LLMs into Jev-style deciders robust to option order, plus self-distilled Tacit models.
ModelresearchPython