Nimble
A local large language model positioned for lightweight, fast-response use cases. Its design emphasizes low latency and rapid inference, making it suitable for scenarios sensitive to response speed. The model can be accessed and run locally through the Ollama platform.
Timeline (last 90 days)
The study compared nine systems on 739 real Italian/English emails (231 unbiased samples and 500 adversarial samples), including keyword rules, TF-IDF+logistic regression, local typed decision models (Tev1, Nimble), a hosted decision model (Nimble v4-max), and hosted generative LLMs (Qwen2.5-7B, Gemma 4 26B/31B).
The Nimble model took 1.7 seconds after warm-up on the email classification task and 1.336 seconds on the scoring task.
After replacing decision nodes with constrained-output models like Nimble, the interface becomes enumerations or scores, and scripts can directly branch or compare thresholds without parsing free text.
The original material did not mention whether Nimble has built-in calibration, so before actual integration it is best to first plot a reliability diagram with known samples to confirm whether the probability distribution is usable.
Nimble is a 9B-class local model.
The --ask mode supports three question forms: yes/no questions, label questions, and rating questions.
The Nimble model can be run in decision mode with the `ollama run nimble --ask` command.
Ollama introduced the experimental decision model Nimble in version 0.40.4.
Bespoke Labs' internal evaluation shows that Nimble reduces the error rate by 43% compared with Jev in actual decision-making scenarios.
Bespoke Labs' internal evaluation shows that Nimble improves accuracy by 50 percentage points compared with Jev on the model routing decision task.
23 more timeline events
All Facts (20)
The study compared nine systems on 739 real Italian/English emails (231 unbiased samples and 500 adversarial samples), including keyword rules, TF-IDF+logistic regression, local typed decision models (Tev1, Nimble), a hosted decision model (Nimble v4-max), and hosted generative LLMs (Qwen2.5-7B, Gemma 4 26B/31B).
50%UnverifiedThe Nimble model took 1.7 seconds after warm-up on the email classification task and 1.336 seconds on the scoring task.
50%UnverifiedThe original material did not mention whether Nimble has built-in calibration, so before actual integration it is best to first plot a reliability diagram with known samples to confirm whether the probability distribution is usable.
50%UnverifiedNimble is a 9B-class local model.
50%UnverifiedThe --ask mode supports three question forms: yes/no questions, label questions, and rating questions.
50%UnverifiedAfter replacing decision nodes with constrained-output models like Nimble, the interface becomes enumerations or scores, and scripts can directly branch or compare thresholds without parsing free text.
50%UnverifiedThe Nimble model can be run in decision mode with the `ollama run nimble --ask` command.
50%UnverifiedSomo is a math tutoring app that takes a photo upload of a handwritten solution process, uses AI to find the first line in the handwritten solution where an error occurs, circles it in red like a teacher, and explains it in English, Swahili, or Spanish.
50%UnverifiedNimble searches the live web for past exam papers on that concept, prioritizing official sites; US students get the NY Regents question bank, and when official sites yield no results it searches the open web; Nimble correctly found the Tanzania Examinations Council's official mathematics papers and examiner reports from 2017 to 2023.
50%UnverifiedPDF-type results return only a summary of about 300–700 characters, which contains no problems; Somo treats Nimble as a discoverer: when it encounters short PDF results, it downloads the files and uses pypdf to extract up to 20,000 characters (processing five papers in parallel), then Gemini picks out a problem whose original text appears verbatim and which tests the same concept.
50%UnverifiedAll calls to Gemini, Gemma, Claude, and Nimble are Sentry spans, and background tasks each form their own transaction; every Gemini-to-Claude fallback also raises a warning issue; Sentry is configured with send_default_pii=False, so student photos and prompt text are not sent.
50%UnverifiedThe Jeff model is a decision-specific model series developed in-house by Bespoke Labs, with parameter counts ranging from 0.8B to 1.4B, while the Nimble series is fine-tuned on a Qwen base.
50%UnverifiedIn Decision Models, each question is scored independently and answers are not passed to subsequent questions; the request body is limited to 64 KB, and Nimble has a context of 8192 tokens.
50%UnverifiedUsing Decision Models requires upgrading to version 0.3.5 or higher, pulling the Nimble model, and sending a POST request to the /decision endpoint; no API key is required for local requests.
50%UnverifiedOn Bespoke Labs' leakage-level evaluation, Nimble reaches 90.12, while its base model Qwen 3.5 9B scores only 66.36.
50%UnverifiedDecision-class models such as Nimble, tev1, clef, and clef-flash have begun to offer support on MLX.
50%UnverifiedThe three typical application scenarios officially listed for Nimble are ticket triage, model routing, and content moderation.
50%UnverifiedNimble is a decision model in Ollama's official examples and can be pulled via the command ollama pull nimble.
50%UnverifiedNimble is positioned as a lightweight, fast-response model.
50%UnverifiedNimble (Nimbleway) has an established background in the web data collection space.
50%