Instructions to use osmapi/osmKeye-VL-2.0-30B-A3B-uncensored with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use osmapi/osmKeye-VL-2.0-30B-A3B-uncensored with Transformers:

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

pipe = pipeline("image-text-to-text", model="osmapi/osmKeye-VL-2.0-30B-A3B-uncensored", trust_remote_code=True)
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 AutoModel
model = AutoModel.from_pretrained("osmapi/osmKeye-VL-2.0-30B-A3B-uncensored", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use osmapi/osmKeye-VL-2.0-30B-A3B-uncensored with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "osmapi/osmKeye-VL-2.0-30B-A3B-uncensored"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "osmapi/osmKeye-VL-2.0-30B-A3B-uncensored",
		"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/osmapi/osmKeye-VL-2.0-30B-A3B-uncensored

SGLang

How to use osmapi/osmKeye-VL-2.0-30B-A3B-uncensored 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 "osmapi/osmKeye-VL-2.0-30B-A3B-uncensored" \
    --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": "osmapi/osmKeye-VL-2.0-30B-A3B-uncensored",
		"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 "osmapi/osmKeye-VL-2.0-30B-A3B-uncensored" \
        --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": "osmapi/osmKeye-VL-2.0-30B-A3B-uncensored",
		"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 osmapi/osmKeye-VL-2.0-30B-A3B-uncensored with Docker Model Runner:
```
docker model run hf.co/osmapi/osmKeye-VL-2.0-30B-A3B-uncensored
```

Keye-VL-2.0-30B-A3B — Abliterated

An abliterated (refusal-direction-ablated) build of Kwai-Keye/Keye-VL-2.0-30B-A3B, a 30B-A3B vision-language MoE. The vision capability is fully preserved — only the text decoder's attention output projections were modified.

⚠️ Use a CUDA stack (vLLM / SGLang / Transformers + flash-attn). This model uses a custom sparse-attention indexer (SALightningIndexer) built for CUDA kernels (flash-attn, DeepGEMM). On Apple Silicon / MPS, generation via Transformers is unstable for the base model too (not an artifact of abliteration) — for Mac, use the MLX builds in the osmapi collection, which run dense attention.

What was changed

Method: Heretic v1.3.0 — automated directional ablation (TPE-optimized over refusal count + KL divergence), 48 trials.
Target: attention o_proj of the text decoder, in the mid/late layers where the refusal direction lives (early layers, lm_head, embed_tokens untouched).
Best trial: marker-refusals 43→37/48, KL ≈ 0.026 (low → general capability largely retained). This is a moderate abliteration: on this batched-MoE architecture Heretic can only reach o_proj (the experts' FFN down_proj are fused batched tensors it doesn't target), so the MLP path's refusal contribution is not ablated. A stronger custom ablation is possible.

Preserved (verified byte-identical to base)

Vision tower visual.vision_model.* (SigLIP encoder + native-resolution packing) — 438 tensors
Projector mlp_AR.* (2×2 spatial-merge → text space)
lm_head, embed_tokens
MTP: the base model has no multi-token-prediction module (nothing to preserve).

Usage (CUDA)

from transformers import pipeline
pipe = pipeline("image-text-to-text",
                model="osmapi/osmKeye-VL-2.0-30B-A3B-uncensored",
                trust_remote_code=True)

Or serve with vLLM / SGLang exactly as the base model.

Provenance / reproducibility

Produced on an Apple M4 Max (128 GB) by abliterating in an isolated venv pinned to transformers==4.57.3 (the model's build version; 5.x removed OutputRecorder). Running this CUDA-built model on macOS required patching out a flash-attn assert, replacing the CUDA fast_hadamard_transform with a pure-torch Walsh–Hadamard transform, and adding AutoModelForImageTextToText to auto_map. Heretic targeted o_proj via a row-normalized LoRA merged into the weights.

Abliteration removes safety alignment; you are responsible for how you use this model.

Other variants of this model (public on `osmapi`)

osmKeye-VL-2.0-30B-A3B-uncensored-mlx-mxfp4 — MLX MXFP4 (~4.43 bpw) — Apple Silicon
osmKeye-VL-2.0-30B-A3B-uncensored-mlx-optiq-3.7bpw — MLX mixed ~3.87 bpw — Apple Silicon

Downloads last month: 26

Safetensors

Model size

31B params

Tensor type

F32

BF16

Model tree for osmapi/osmKeye-VL-2.0-30B-A3B-uncensored

Base model

Kwai-Keye/Keye-VL-2.0-30B-A3B

Finetuned

(1)

this model