Bespoke Nimble
Bespoke Labs' open recipe, data and 9B LoRA model for Jev-style choice and true/false decisions, runnable on Apple Silicon or NVIDIA GPUs.
Overview
Bespoke Nimble shares data, a model and a recipe for an open Jev-style decision model. Nimble takes text and a schema of questions, each a choice from a list or a true/false question, and returns the picked answer and the probability of each allowed answer in one step without writing reasoning. The authors stress they did not distill from Jev; the goal is to show how to curate data, train and serve such a model.
Bespoke-Nimble-9B is a LoRA on Qwen3.5-9B trained on contrastively curated synthetic data; the latest checkpoint has an 8,192-token context and up to 255 choices per field. It runs on Apple Silicon or NVIDIA GPUs and is available in Ollama as nimble. A public 13-dataset benchmark suite can run Nimble and Jev side by side; list summaries report 90.1% on a 324-example holdout against Jev's 93.2%. The data pipeline, training config and eval harness are Apache-2.0.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which listed option fits the text? | — |
| Yes / no | Is this statement about the text true? | true · false |
How it works
- Curate contrastive synthetic data
- Fine-tune a LoRA on Qwen3.5-9B
- Serve locally (Mac or NVIDIA, or Ollama)
- Benchmark against Jev on the public suite
Strengths
- Full open recipe including data pipeline
- Runs locally
- Side-by-side benchmark with Jev
Caveats
- Slightly behind Jev on its own holdout per listing
- 9B model needs a capable machine
Getting started
Download Bespoke-Nimble-9B from Hugging Face or run it in Ollama as nimble; see the Quickstart.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- local model (Apple Silicon / NVIDIA / Ollama)
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Open alternative: a LoRA on Qwen3.5-9B that scores one allowed answer token per `Choice`, boolean or rubric-score question, released with its data pipeline, training config and eval harness under Apache-2.0, and reporting 90.1% on its 324-example holdout against Jev's 93.2%.
yibie/awesome-jevBespoke Labs' 9B LoRA fine-tune of Qwen3.5-9B on contrastively curated synthetic data, with a public 13-dataset benchmark suite. In Ollama as `nimble`.
AnotiaWang/awesome-decision-modelsOpen data, training recipe, and a 9B model for Jev-style choice and true/false decisions on Apple Silicon or NVIDIA GPUs.
MrJev/awesome-jev
Frequently asked questions
Was it distilled from Jev?
No, the authors say they did not distill from Jev.
How does it compare?
Listing reports 90.1% on a 324-example holdout versus Jev's 93.2%.
Where can I run it?
On Apple Silicon or NVIDIA GPUs, or via Ollama.
Related projects
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython