Instructions to use BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B with Transformers:

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

pipe = pipeline("text-generation", model="BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B")
model = AutoModelForCausalLM.from_pretrained("BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B

SGLang

How to use BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B 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 "BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B with Docker Model Runner:
```
docker model run hf.co/BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B
```

DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B

Expert-pruned variant of BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8, a mixed-precision quantization of deepseek-ai/DeepSeek-V4-Flash-0731.

What changed: routed experts pruned uniformly from 256 → 192 per layer across all 43 MoE layers. Quantization formats, dense layers, and KV cache are unchanged from the source — see its card for quantization and calibration details.

How K=192 was selected: per-expert REAP saliency computed during INT4 calibration,

0.7 × num_captured × (1 − rel_l2_error) + 0.3 × gate_l2_norm

keeping the top-192 experts in every layer. The per-layer mask is stored in reap_saliency.json; surviving experts are compacted to contiguous indices 0..191.

Parameters (INT4 experts unpacked ×8):

Component	Params
Routed experts (192 × 43 × 25.17 M)	207.8 B
Dense (attention, shared experts, indexer, norms, embeddings)	7.8 B
Total weights (≈ 216 B)	215.6 B
Active per token (6 routed + shared + dense)	≈ 12 B

Quantization scales and index tables add ≈ 6.6 B of storage tensors (not counted as model parameters). Before pruning (K=256) the model is ≈ 285 B.

Usage:

vllm serve BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B \
  --trust-remote-code --kv-cache-dtype fp8

License: MIT — same as deepseek-ai/DeepSeek-V4-Flash-0731.

Cite:

@misc{deepseekai2026deepseekv4,
      title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
      author={DeepSeek-AI},
      year={2026},
}

Downloads last month: 84

Safetensors

Model size

216B params

Tensor type

I64

F32

I32

BF16

F8_E4M3

Model tree for BlivionIaG/DeepSeek-V4-Flash-0731-Int4-FP8-REAP-216B

Base model

deepseek-ai/DeepSeek-V4-Flash-0731

Quantized

(192)

this model