Nimble: Complete Guide to Installation and Use with Ollama

Nimble is a 9-billion-parameter decision model from Bespoke Labs, fine-tuned from Qwen3.5-9B. It is not designed as a general-purpose chat model. Instead of generating long free-form answers, Nimble takes text or structured state plus a set of predefined questions, selects an allowed answer for each question, and returns probability information for the available choices.

Primary source: Nimble in the official Ollama Model Library. Because model tags and runtime behavior can change, that page should be treated as the main source of truth for current Ollama usage.

What is Nimble for?

Nimble is useful when an application needs a decision, not another paragraph of generated text. Typical uses include:

• support-ticket routing;
• boolean condition checks;
• policy-compliance checks before an AI agent executes a tool;
• rubric-based scoring;
• model or agent routing;
• CRM lead and request classification;
• fast guardrails for tool-using agents.

This makes it particularly relevant to agentic AI, CRM systems, workflow automation, and tool-using agents.

How is it different from a normal LLM?

General chat models generate tokens to form an open-ended answer. Nimble’s evaluated decision workflow directly scores the allowed answer tokens instead of first generating a long textual reasoning trace.

This gives applications a predictable, typed result that is easier to consume in backend services and automation pipelines.

Ollama reports a sub-100 ms result in a demonstration on a MacBook Pro equipped with an M5 Max. This is a hardware-specific measurement, not a universal latency guarantee.

Key specifications

• Developer: Bespoke Labs
• Base model: Qwen3.5-9B
• Size: approximately 9B parameters
• Input: text
• Primary task: typed decisions / classification
• License: Apache 2.0
• Required Ollama version: 0.35 or later
• Practical decision-workflow context limit: approximately 8,192 tokens
• Questions through /v1/systemone: 1 to 64 per request
• Main question types: choice, noul, and score

Why does an Ollama tag show 256K context?

Some Ollama tag metadata displays a 256K context window, while the current Nimble checkpoint and its decision workflow document an 8,192-token limit.

For production use, design Nimble prompts around the 8K decision limit and keep context and schemas concise.

Question types

Choice

Select one option from a list.

Noul

The current Ollama Decision API uses noul for a boolean-style True/False decision.

Score

Place the input on an ordered rubric.

For the current /v1/systemone API, Ollama documents 2–26 options for choice and score. The underlying model format can support larger schemas, but applications using System One should follow the API’s current limits.

Requirements

Install Ollama first.

• Windows 10 or later
• macOS: the current official download page lists macOS 14 Sonoma or later
• Linux: use Ollama’s official Linux or manual installation instructions
• A GPU or Apple Silicon is recommended for comfortable 9B-model performance
• RAM/VRAM requirements depend on quantization and context length

Install Ollama

Windows

Open the official Ollama download page for Windows. You can use the installer or run the official PowerShell installer:

irm https://ollama.com/install.ps1 | iex

Verify the installation:

ollama --version

Nimble requires Ollama 0.35 or later.

macOS

Download Ollama from the official macOS page or use the official Terminal installer:

curl -fsSL https://ollama.com/install.sh | sh

Then verify:

ollama --version

Linux

See the official Linux installation page. The official quick-install command is:

curl -fsSL https://ollama.com/install.sh | sh

Then verify:

ollama --version

Download and run Nimble

Official model page: Nimble on Ollama

Pull the default model:

ollama pull nimble

Run it:

ollama run nimble

List locally installed models:

ollama list

See every available variant on the Nimble tags page.

Available model variants

At publication time, the main Ollama tags include:

A lighter starting point for many local machines is:

ollama pull nimble:9b-q4_K_M

If memory is available, Q8 or BF16 can also be evaluated.

Is ollama run nimble enough?

It is useful for loading or basic testing, but Nimble’s primary typed-decision workflow uses Ollama’s dedicated decision endpoint:

localhost:11434/v1/systemone

Ollama’s current documentation notes that decision models are not yet exposed through every normal feature of the Ollama Python and JavaScript libraries. For the intended workflow, use the REST API or TypeSafe SDK.

Understanding /v1/systemone

A request contains several important fields.

model

"model": "nimble"

state

The information Nimble must judge. It can be a string, object, or array.

questions

One or more typed questions about the state.

keep_alive

An optional value controlling how long the model remains loaded.

First cURL example

Imagine a CRM system that must route a customer message and determine whether a refund is requested:

curl -s localhost:11434/v1/systemone \
 -H "Content-Type: application/json" \
 -d '{
 "model": "nimble",
 "state": {
 "message": "My card was charged twice and I need the extra payment reversed."
 },
 "questions": {
 "department": {
 "type": "choice",
 "instructions": "Choose the best team for this request.",
 "criteria": {
 "billing": "Payments, invoices and refunds",
 "technical": "Product bugs and integrations",
 "sales": "Plans and purchases"
 }
 },
 "needs_refund": {
 "type": "noul",
 "instructions": "Does the user request a refund or reversal?"
 },
 "priority": {
 "type": "score",
 "instructions": "Rate how quickly this should be reviewed.",
 "criteria": ["Routine", "Soon", "Urgent"]
 }
 }
 }'

The response includes the selected answer for each question and probability information for allowed outcomes.

What does probability mean?

Nimble does not only return a label. It can expose a probability distribution over the answers you defined.

That enables workflows such as:

High-confidence result → automatic action
Uncertain result → human review

Thresholds should be calibrated using your own data rather than chosen arbitrarily.

What does confidence mean?

confidence describes how concentrated the probability distribution is.

It is not a direct guarantee of correctness.

A probability of 0.90 does not automatically mean 90% accuracy on your production data. Evaluate the model on data from your own domain.

Schema design tips

1. Make options clear and as mutually exclusive as possible.
2. Add other or no_match when none of the listed choices may fit.
3. Keep context short and relevant.
4. Write criteria explicitly.
5. Include policies clearly in the state or instructions.
6. Calibrate thresholds on realistic data.
7. Measure false positives and false negatives before production.

Questions are scored independently

If two answers need to remain logically consistent, validate that relationship in your own application code.

Nimble should not be treated as a replacement for deterministic business rules where those rules can be implemented directly.

Using Nimble as an Agent Guardrail

A useful architecture is:

User Request
 ↓
LLM Agent
 ↓
Proposed Tool Call
 ↓
Nimble Decision Layer
 ↓
Policy / Risk / Approval Check
 ↓
Execute Tool OR Block OR Human Review

This prevents a generative model from being the only component deciding whether an important tool call should execute.

Tool-call moderation

You can place a proposed tool call in state and ask questions such as:

• Could this tool call cause harm?
• Does it require human approval?
• Is it compliant with organizational policy?
• What is its risk level?

Nimble can then operate as a fast decision layer before execution.

CRM use cases

Ticket Routing

Choose the team that should receive a support request.

Lead Classification

Classify a lead using structured context.

Refund Eligibility

Check whether a request meets predefined refund criteria.

Priority Scoring

Assign a ticket or opportunity to an urgency rubric.

Escalation Detection

Decide whether a case should be handed to a manager or human agent.

Policy Enforcement

Evaluate an agent’s proposed action against policy.

Human Handoff

Determine whether the conversation should leave the automated flow.

Model Routing

Choose the next model or agent for a request.

Does Nimble replace a generative LLM?

Usually not.

A practical architecture can combine:

Generative LLM → content and open-ended reasoning
Nimble → constrained typed decisions
Rules Engine → deterministic rules
Human Review → sensitive or uncertain cases

Nimble is best understood as a decision layer, not a chatbot replacement.

What Nimble is not designed for

• long-form article generation;
• creative writing;
• open-ended summarization;
• chain-of-thought generation;
• arbitrary quote extraction;
• arbitrary nested JSON generation;
• open-ended question answering.

Nimble chooses among answers that you define.

Contrastive training

Bespoke Labs trained Nimble using contrastive examples where two samples are very similar but one important fact changes. That fact is expected to flip the correct decision.

The objective is to teach the model which evidence should actually change a decision.

Evaluation

Ollama and Bespoke Labs publish public benchmark results for Nimble. Ollama reports an evaluation spanning 13 public datasets and 3,880 human-labeled decisions across choice, boolean, and score tasks.

Benchmarks demonstrate capability, but they are not a guarantee of performance on your organization’s data.

Privacy and local execution

When Nimble is run locally through Ollama, inference is performed on your own machine.

That is valuable for workloads involving:

• private CRM records;
• customer support tickets;
• internal organizational data;
• agent tool calls;
• private policy documents.

Overall privacy still depends on any external APIs or services used elsewhere in the architecture.

List installed models

ollama list

Remove Nimble

ollama rm nimble

Update Nimble

ollama pull nimble

Production recommendation

Using nimble or latest is convenient during development. For reproducible production deployments, pin an explicit tag such as:

nimble:9b-q4_K_M

or:

nimble:9b-q8_0

Evaluate each update before rolling it out.

Quick-start checklist

1. Install Ollama
2. Verify Ollama >= 0.35
3. Choose a Nimble quantization
4. Run ollama pull nimble
5. Test /v1/systemone
6. Define Choice / Noul / Score questions
7. Evaluate on your own data
8. Calibrate thresholds
9. Add human review for sensitive cases
10. Pin a specific model tag in production

Official links and downloads

Use the official links below for installation, model versions, and current documentation:

• Nimble on the Ollama Model Library — model overview, usage, and /v1/systemone documentation
• All Nimble tags and variants — Q4, Q8, BF16, and current download sizes
• Download Ollama for Windows
• Download Ollama for macOS
• Install Ollama on Linux
• Official Nimble repository on GitHub
• Bespoke-Nimble-9B on Hugging Face
• Ollama blog: Jev-style Decision Models

Conclusion

Nimble is not a general chatbot. It is a fast, typed decision model designed for routing, boolean checks, rubric scoring, policy enforcement, tool moderation, human handoff, and model routing.

To get started, install Ollama 0.35 or later, run ollama pull nimble, and build the production decision flow around /v1/systemone.