Instructions to use bespokelabs/Nimble-v4-max with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use bespokelabs/Nimble-v4-max with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="bespokelabs/Nimble-v4-max")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("bespokelabs/Nimble-v4-max")
model = AutoModelForMultimodalLM.from_pretrained("bespokelabs/Nimble-v4-max", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use bespokelabs/Nimble-v4-max with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "bespokelabs/Nimble-v4-max"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bespokelabs/Nimble-v4-max",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/bespokelabs/Nimble-v4-max

SGLang

How to use bespokelabs/Nimble-v4-max with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "bespokelabs/Nimble-v4-max" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bespokelabs/Nimble-v4-max",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "bespokelabs/Nimble-v4-max" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bespokelabs/Nimble-v4-max",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use bespokelabs/Nimble-v4-max with Docker Model Runner:
```
docker model run hf.co/bespokelabs/Nimble-v4-max
```

Nimble-v4-max

Nimble-v4-max is a decision model made by Bespoke Labs, a fine-tune of Qwen/Qwen3.8-27B. Given a context and typed questions (a choice among options, a yes or no, or a score on ordered levels), it returns a probability for each option.

This is version 4.1. It scores 72.74 on the Decision Index 0.3 public suite, up from 68.46 for v4.0, with the largest gains on intent classification, math, reasoning and select-all questions. It takes the same requests and runs with the same inference code as v4.0, so it is a drop-in replacement. The full run, the inference code and instructions are in bespokelabs/nimble-v4-max-decision-index.

Versions

Version	Revision	Decision Index 0.3 public	Default temperature
v4.1 (2026-11)	tag `v4.1`	72.74	2.25
v4.0 (2026-10)	tag `v4.0` (`655aa705`)	68.46	1.5

To keep v4.0, pin revision="v4.0".

Results on well-known benchmarks

From the Decision Index 0.3 public run, on the benchmarks' own metrics. Scores for all 43 benchmarks are in benchmark-summary.json in the results dataset.

Benchmark	Metric	v4.0	v4.1
MMLU	accuracy	87.9	88.3
MMLU-Pro	accuracy	71.0	74.1
GPQA Diamond	accuracy	52.5	55.6
GSM8K	accuracy	75.9	83.6
HLE	accuracy	12.4	15.2
BBH (fixed-option tasks)	accuracy	93.5	93.8
MuSR	accuracy	70.0	75.7
ARC-Challenge	accuracy	97.0	97.1
HellaSwag	accuracy	97.9	97.7
WinoGrande	accuracy	94.2	94.5
CRUXEval	accuracy	86.7	86.5
CLadder	accuracy	89.5	92.9
ANLI	macro-F1	75.5	78.7
BANKING77	macro-F1	79.9	93.2
CLINC150 + OOS	macro-F1	91.6	96.2
BFCL	case exact accuracy	98.0	98.3
When2Call	accuracy	90.7	91.4
ToolRet	nDCG@10	72.9	73.3
BRIGHT	nDCG@10	49.7	48.6
Amazon ESCI	macro-F1	62.9	62.1
RAGTruth	F1, hallucinated class	86.0	86.1

How to run it

Use the inference code in bespokelabs/nimble-v4-max-decision-index (code/nimble/). It builds the prompts that schema_config.json pins by hash, reads each request with the attention pattern the model was trained with, and scores one token per option. Loading the weights as a plain chat model, or generating text, does not reproduce its answers.

The inference code returns the softmax over the option logits at temperature 1. For calibrated probabilities, divide the option logits by the default temperature in the table above before the softmax. Temperature never changes which option ranks first.

The model needs about 55 GB in BF16, so it runs on one 80 GB or larger GPU.

License

Nimble-v4-max is released under CC BY-NC 4.0, so it may not be used for commercial purposes. The base model, Qwen/Qwen3.8-27B, is licensed separately under Apache 2.0; its license text is in LICENSE-Qwen3.8-27B.

Citation

@misc{bespokelabs2026nimblev4max,
  title        = {Nimble-v4-max},
  author       = {Shreyas Pimpalgaonkar and Anirudha Ramesh and Sid Devic and Mrinal Kumar and Pratik Dalal and Mahesh Sathiamoorthy},
  year         = {2026},
  publisher    = {Bespoke Labs},
  howpublished = {\url{https://huggingface.co/bespokelabs/Nimble-v4-max}}
}

Downloads last month: 227

Safetensors

Model size

27B params

Tensor type

BF16

Model tree for bespokelabs/Nimble-v4-max

Base model

Qwen/Qwen3.8-27B

Finetuned

(830)

this model

bespokelabs
/

Nimble-v4-max