Instructions to use truespirit/Ornith-1.5-397B-MXFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use truespirit/Ornith-1.5-397B-MXFP4 with Transformers:

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

pipe = pipeline("text-generation", model="truespirit/Ornith-1.5-397B-MXFP4")
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("truespirit/Ornith-1.5-397B-MXFP4")
model = AutoModelForMultimodalLM.from_pretrained("truespirit/Ornith-1.5-397B-MXFP4", 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 truespirit/Ornith-1.5-397B-MXFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "truespirit/Ornith-1.5-397B-MXFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "truespirit/Ornith-1.5-397B-MXFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/truespirit/Ornith-1.5-397B-MXFP4

SGLang

How to use truespirit/Ornith-1.5-397B-MXFP4 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 "truespirit/Ornith-1.5-397B-MXFP4" \
    --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": "truespirit/Ornith-1.5-397B-MXFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "truespirit/Ornith-1.5-397B-MXFP4" \
        --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": "truespirit/Ornith-1.5-397B-MXFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use truespirit/Ornith-1.5-397B-MXFP4 with Docker Model Runner:
```
docker model run hf.co/truespirit/Ornith-1.5-397B-MXFP4
```

Ornith-1.5-397B-MXFP4

MXFP4-quantized build of ornith-ai/Ornith-1.5-397B — the same model at roughly half the disk and memory footprint, targeting single-GPU serving on 288 GB AMD Instinct MI355X.

Quantization


Source	`ornith-ai/Ornith-1.5-397B-FP8` (FP8, compressed-tensors)
Tool	AMD Quark 0.12.post1, file-to-file
Recipe	AMD's published Qwen3.5-397B-A17B-MXFP4 recipe
Scheme	OCP MXFP4 weights, block size 32 (stored as packed U8 + E8M0 scales — hence the BF16·U8 tensor listing)
Excluded (kept BF16)	`lm_head`, visual tower, MTP, MoE gates, shared experts, attention (incl. linear attention)
Size	239 GB (122 shards) vs ~790 GB BF16

Only the routed MoE experts carry MXFP4 weights (*.mlp.experts.* with weight_scale); every sensitive or non-MoE component is untouched.

Usage

Requires an inference stack with OCP MXFP4 checkpoint support — e.g. vLLM built for ROCm with AMD Quark enabled. TP=1 load fits a single MI355X (288 GB).

vllm serve truespirit/Ornith-1.5-397B-MXFP4 --tensor-parallel-size 1

Notes

Structure, shard integrity, tensor formats, and MXFP4 scale presence are validated; generation quality vs. the original is not benchmarked yet.
Quantization can introduce small differences in numerical behavior and outputs. The config reports model_type=qwen3_5_moe (Qwen3_5MoeForConditionalGeneration), so architecture details, intended use, and limitations follow the original model card.

License

MIT — inherited from the base model. Review ornith-ai/Ornith-1.5-397B before use.

Downloads last month: -

Safetensors

Model size

210B params

Tensor type

BF16

Model tree for truespirit/Ornith-1.5-397B-MXFP4

Base model

ornith-ai/Ornith-1.5-397B

Quantized

(8)

this model