Instructions to use AtomicChat/Ling-3.0-flash-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AtomicChat/Ling-3.0-flash-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="AtomicChat/Ling-3.0-flash-NVFP4", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("AtomicChat/Ling-3.0-flash-NVFP4", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AtomicChat/Ling-3.0-flash-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AtomicChat/Ling-3.0-flash-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AtomicChat/Ling-3.0-flash-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AtomicChat/Ling-3.0-flash-NVFP4

SGLang

How to use AtomicChat/Ling-3.0-flash-NVFP4 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 "AtomicChat/Ling-3.0-flash-NVFP4" \
    --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": "AtomicChat/Ling-3.0-flash-NVFP4",
		"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 "AtomicChat/Ling-3.0-flash-NVFP4" \
        --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": "AtomicChat/Ling-3.0-flash-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AtomicChat/Ling-3.0-flash-NVFP4 with Docker Model Runner:
```
docker model run hf.co/AtomicChat/Ling-3.0-flash-NVFP4
```

worthant commited on Aug 5

Commit

91e518c

verified ·

1 Parent(s): b0ff5db

docs: model card

Browse files

Files changed (1) hide show

README.md +48 -0

README.md ADDED Viewed

	@@ -0,0 +1,48 @@

+---
+license: mit
+base_model: inclusionAI/Ling-3.0-flash
+base_model_relation: quantized
+pipeline_tag: text-generation
+tags:
+  - gguf
+  - nvfp4
+  - blackwell
+  - llama.cpp
+library_name: llama.cpp
+---
+# Ling-3.0-flash — NVFP4
+[`inclusionAI/Ling-3.0-flash`](https://huggingface.co/inclusionAI/Ling-3.0-flash) (124B total, 5.1B active) with the routed experts in NVFP4.
+NVFP4 packs 64 weights into 36 bytes: 4-bit E2M1 values with a UE4M3 scale for every 16 of them. The scale granularity is twice as fine as MXFP4, which is what makes it usable at this bit width.
+| | |
+| --- | --- |
+| size | 72.3 GB |
+| experts | NVFP4 (4.50 bpw) |
+| attention, KDA gates, shared experts | Q8_0 |
+| router, norms, `A_log`, `dt_bias`, conv1d | F32 |
+| output | F16 |
+Fits a single 128 GB DGX Spark with room for a long context.
+## Requirements
+The architecture is not in upstream llama.cpp yet:
+```bash
+git clone -b feat/bailingmoe3 https://github.com/AtomicBot-ai/atomic-llama-cpp-turboquant llama.cpp
+cd llama.cpp && cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="120;121" && cmake --build build -j
+llama-cli -m Ling-3.0-flash-NVFP4-00001-of-00002.gguf --jinja -ngl 99 -c 32768
+```
+**Blackwell is what this format is for.** The native FP4 tensor cores exist on sm_100 (B200) and sm_120 / sm_121 (RTX PRO 6000, RTX 5090, DGX Spark). On older cards the file still loads and runs through the dequantization path, just without the speedup.
+## Notes on the build
+The importance matrix is not used here: the NVFP4 quantizer in llama.cpp ignores `quant_weights` and rounds to nearest inside each 16-element block. Bits were spent where they matter instead — everything outside the routed experts is kept at Q8_0 or better, which costs about 4 GB.
+Measurements against the BF16 baseline, and the same numbers for the whole K/IQ grid, are in [`AtomicChat/Ling-3.0-flash-GGUF-metrics`](https://huggingface.co/datasets/AtomicChat/Ling-3.0-flash-GGUF-metrics).
+<!-- RESULTS -->