Instructions to use kirp/jpt-9b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kirp/jpt-9b with Transformers:

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

pipe = pipeline("image-text-to-text", model="kirp/jpt-9b")
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)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("kirp/jpt-9b")
model = AutoModelForMultimodalLM.from_pretrained("kirp/jpt-9b", 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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use kirp/jpt-9b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kirp/jpt-9b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "kirp/jpt-9b",
		"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/kirp/jpt-9b

SGLang

How to use kirp/jpt-9b 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 "kirp/jpt-9b" \
    --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": "kirp/jpt-9b",
		"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 "kirp/jpt-9b" \
        --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": "kirp/jpt-9b",
		"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 kirp/jpt-9b with Docker Model Runner:
```
docker model run hf.co/kirp/jpt-9b
```

JPT-9B

The large sibling of JPT-4B: the same recipe and the same data (mix_train_env_v11) on Qwen/Qwen3.5-9B. Give it a situation (text, optionally with images) and typed questions — choice, score or noul (yes/no). It returns a probability for every option from one forward pass, with no generated text and no reasoning tokens.

Results

JevBench public items (231, frozen since v1.2), our own runner (same harness as the JPT-4B / JPT-0.8B rows; only the maintainer can run the sealed 308):

System	Params	Public accuracy (231)	Source
JPT-9B	9B	0.857	ours
Jev 1.13.0 (TypeSafe AI, API)	closed	0.866	JevBench v1.4 results
JPT-4B	4B	0.879	ours
JevK5 v0.2.0	27B	0.853	JevBench v1.4 results
Winnow-12B Q8	12B	0.857	JevBench v1.4 results
decider-35b-a3b	35B-A3B	0.831	JevBench v1.4 results
Hopper	—	0.823	JevBench v1.4 results
openjev 4B v5 (AlexWortega)	4B	0.814	its own card
SemIf, formerly OpenJev (TheoLeeCJ, Qwen3.5-4B)	4B	0.810	JevBench v1.4 results
local-jev Qwen3.5-4B	4B	0.805	JevBench v1.4 results
reflex 4B	4B	0.792	JevBench v1.4 results
kev 4B (research preview)	4B	0.662	JevBench v1.4 results

JPT-9B has no official v1.4 score yet, and we have not yet run the NeoHorse-Jev-4B head-to-head comparison or the image evals (ScreenSpot-v2, Screen2Words, ERQA) on this checkpoint — see the JPT-4B card for those.

Other benchmarks, at the fitted temperature T = 1.087 (JPT-4B's T = 1.036 shown for context):

Benchmark	JPT-9B	JPT-4B
JevBench public hard tier (111)	0.730 (ECE 0.097, Brier 0.382)	0.784
Typed decisions test (2,000, in-distribution)	0.806 (ECE 0.149, Brier 0.316)	0.796
ANLI r1 / r2 / r3 (dev)	0.737 / 0.647 / 0.677	0.697 / — / 0.613
Banking77 / MASSIVE en / de / zh (in-distribution)	0.787 / 0.897 / 0.853 / 0.830	0.757 / 0.857 / 0.833 / 0.837
AG News / Emotion / SST-5 (out-of-distribution)	0.917 / 0.557 / 0.603	—
EnvBench v0.1 public / held-out (skill, 0–100)	50.5 / 46.8	47.7 / 47.0
Qwen3.5-9B, same prompt, zero-shot — EnvBench public / held-out	24.2 / 25.3	—

ood_massive_zh is a near-tie with JPT-4B (0.830 vs 0.837). jevbench_hard and the EnvBench held-out game area (0.288 vs 0.315) are the only two spots JPT-9B does not improve on JPT-4B — see Limitations.

Decision Index

Decision Index 0.2 (2026-09-24), the full frozen suite (162,841 requests, 40 benchmarks, chance-corrected), run ourselves through llm2jev over SGLang and submitted as apolinario/decision-index#7:

Model	Params	Decision Index
JPT-9B	9B	42.73
Decision 1.0 Lux	9B	38.98
Bespoke Nimble 9B v2	9B	36.68
Kev 9B	9B	35.41

Full run and scores.json: kirp/decision-index-results-jpt-9b (gated: it carries the suite's GPQA/HLE item text).

Quick start

The same as JPT-4B. Swap in the model name and temperature:

python -m sglang.launch_server --model-path kirp/jpt-9b --port 30000 \
  --context-length 32768 --mamba-scheduler-strategy extra_buffer &  # Qwen3.5's DeltaNet layers need this flag
llm2jev --model kirp/jpt-9b --backend sglang --url http://127.0.0.1:30000 --port 8080 --temperature 1.087

No GPU / no engine: pip install "llm2jev[hf,vision]", then llm2jev --model kirp/jpt-9b --backend hf --port 8080 --temperature 1.087. Requests go to POST /v1/systemone; examples are on the JPT-4B card.

Training

The recipe matches JPT-4B's (mix_train_env_v11, 49,221 questions): LoRA r=16 on every projection of the language model, merged into full weights; lr 5e-5, a multi-class Brier loss over option labels, llm2jev's chat prompt with thinking off. It trained on 8 GPUs × 1 × 5 gradient-accumulation steps = 40 questions per step (1 epoch over two option-shuffled copies, materialized into the training file). The vision tower is untouched. No item from JevBench, EnvBench's held-out seeds, the Decision Index frozen suite or our typed test split was used in training.

Limitations

jevbench_hard (0.730 vs JPT-4B's 0.784) and the EnvBench held-out game area (0.288 vs 0.315) are the two places this checkpoint does not improve on the smaller JPT-4B, despite a lower training/validation loss throughout. Not yet root-caused; don't assume larger strictly means better on hard multi-step decisions.
It answers from the evidence it is given: no reasoning phase by design, weakest on multi-step arithmetic and date computation.
Accepts up to 255 options; training covered up to 77 (Banking77).
English first. Image questions work zero-shot through the base vision tower (not evaluated on this checkpoint yet).

License

CC BY-NC 4.0. The weights derive from Qwen3.5-9B (Apache-2.0), but some training datasets allow only non-commercial or research use, so the model is released for non-commercial use.

JPT-9B is an independent open model that implements a typed-decision interface (noul, choice and score questions answered with probabilities). It is not affiliated with, endorsed by or derived from TypeSafe AI or its Jev model, and it was not trained on Jev outputs.

Downloads last month: 31

Safetensors

Model size

9B params

Tensor type

F32

BF16

Model tree for kirp/jpt-9b

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Finetuned

(942)

this model

kirp
/

jpt-9b