Nimble: Bespoke Labs' 9B decision model runs in Ollama

What this is about

Bespoke Labs has released Nimble, a decision model with 9 billion parameters (roughly speaking, the tunable settings learned during training). It's available in the Ollama library and was fine-tuned from Qwen3.5-9B. You can start it with ollama run nimble or download it with ollama pull nimble. There's one catch, though: Nimble is not a chatbot. You give it a text and a set of questions. For each question, it picks an answer and reports the probability of every allowed answer.

How it works

Nimble reads the prompt once per question and directly scores the probabilities of the answer tokens (small units of text). There is no thinking or reasoning step. According to the Ollama page, that's exactly what makes the model fast. In the sample response, usage for three questions is just 841 input tokens and 3 output tokens.

Each question has a type, instructions and usually criteria. These are the types:

• choice: Pick from 2 to 26 options, e.g. which team should handle a ticket. The output is the chosen option and the probability of each option.
• score: A scale with 2 to 26 levels, e.g. "Routine, Soon, Urgent". The output is a value between 0 and 2 plus the probabilities for each level.
• noul: A yes/no-style question that returns a value between 0 and 1, for example whether a customer explicitly demands a refund.

The confidence field ranges from 0 to 1 and shows how concentrated the probabilities are. The source stresses that it is not the chance that the answer is correct.

Example and access

In the sample ticket "I was charged twice. Please refund the extra payment.", Nimble assigns the ticket to the "billing" team with 0.984. The refund question gets 0.997, and urgency comes out at 0.693 (with a low confidence of 0.068, since the probabilities are spread out).

Decision models use the Ollama endpoint /v1/systemone, which follows TypeSafe's Jev API. The text to be assessed goes in state, and the questions go in questions. Ollama builds the prompt itself. Alternatively, there's the official Python SDK typesafe-sdk, which you point at Ollama (TYPESAFE_BASE_URL=http://localhost:11434). It requires an API key, which Ollama ignores, so any value works. This kind of model isn't yet supported in the Ollama CLI or in the Python and JavaScript libraries.

Use cases

• Ticket triage: Options can also be given without a description (null), in which case the names "explain themselves".
• Model routing: Each request goes to the right model, such as a small model for simple tasks and a large one for hard tasks like a database schema.
• Vetting tool calls: An agent call like rm -rf ~ can be checked for possible damage before it runs.

Training and benchmarks

Bespoke Labs builds training data in pairs: two examples are identical except for one fact, and when that fact changes, the correct answer flips. This is meant to teach the model which evidence should change a decision. Before an example is admitted, separate model calls check that the facts fit the rule, that the text doesn't give away the answer, and that dropping either of the two evidence sentences leaves the decisive fact open. The labels are produced by running the verified rules in code.

Bespoke Labs held back 324 examples (162 contrast pairs) for testing. The company also ran Nimble and Jev 1.13.0 on the same 3,880 records from 13 public datasets with human labels. None of the tasks belongs to Nimble's training categories. Accuracy means agreement with the human label. For rubric subsets, only an exact match of the level counts, which is why some values come out low for both models. Nimble and Tev1 ran on Ollama, while the Jev result comes from Bespoke Labs' own run through TypeSafe's API. The source at hand gives no concrete numbers, and all claims come from the vendor and apply to the original Nimble version.

What you need to try it

The source gives no hardware figures. As a rough estimate, a 9B model in 4-bit needs about 6 to 8 GB of VRAM (video memory) or matching system RAM. Ollama also runs on Mac and on CPU, though more slowly. Plan for a few GB of disk space.

Our take

Nimble shows a trend: small, specialized models for structured decisions instead of long text answers. For classification, routing and moderation, that's practical, because you get probabilities instead of free text and can set thresholds. On the critical side, the quality claims come from Bespoke Labs itself, and the source gives no concrete numbers. Confidence isn't a measure of correctness. Integration into Ollama's CLI and libraries is also still missing, and the API format depends on TypeSafe's Jev API.