Instructions to use 0xSero/DeepSeek-V4-Flash-162B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use 0xSero/DeepSeek-V4-Flash-162B with Transformers:

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

pipe = pipeline("text-generation", model="0xSero/DeepSeek-V4-Flash-162B")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("0xSero/DeepSeek-V4-Flash-162B")
model = AutoModelForCausalLM.from_pretrained("0xSero/DeepSeek-V4-Flash-162B")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use 0xSero/DeepSeek-V4-Flash-162B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "0xSero/DeepSeek-V4-Flash-162B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "0xSero/DeepSeek-V4-Flash-162B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/0xSero/DeepSeek-V4-Flash-162B

SGLang

How to use 0xSero/DeepSeek-V4-Flash-162B 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 "0xSero/DeepSeek-V4-Flash-162B" \
    --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": "0xSero/DeepSeek-V4-Flash-162B",
		"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 "0xSero/DeepSeek-V4-Flash-162B" \
        --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": "0xSero/DeepSeek-V4-Flash-162B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use 0xSero/DeepSeek-V4-Flash-162B with Docker Model Runner:
```
docker model run hf.co/0xSero/DeepSeek-V4-Flash-162B
```

0xSero commited on 29 days ago

Commit

2b29d03

verified ·

1 Parent(s): d40cb10

Standardize model card (template rollout)

Browse files

Files changed (1) hide show

README.md +56 -11

README.md CHANGED Viewed

@@ -3,19 +3,50 @@ license: mit
 library_name: transformers
 pipeline_tag: text-generation
 tags:
-  - deepseek-v4
-  - mixture-of-experts
-  - reap
-  - dgx-spark
-  - vllm
-  - long-context
-  - fp8
-  - mxfp4
-  - experimental
-base_model: deepseek-ai/DeepSeek-V4-Flash
 ---
-# DeepSeek-V4-Flash-Spark-Mini
 **162B parameters | K144 REAP-pruned | 200K context | no speculative decoding**
@@ -193,3 +224,17 @@ Choose this if you value prefill throughput over decode speed, or if you want a
 ## License
 MIT for the serving recipe and tooling. The base model weights follow the DeepSeek V4 Flash license. Review it before use.

 library_name: transformers
 pipeline_tag: text-generation
 tags:
+- deepseek
+- deepseek-v4
+- dgx-spark
+- experimental
+- fp8
+- long-context
+- mixture-of-experts
+- mxfp4
+- reap
+- vllm
+base_model:
+- deepseek-ai/DeepSeek-V4-Flash
 ---
+> [!TIP]
+> **[Support this work →](https://donate.sybilsolutions.ai)** · [X](https://x.com/0xsero) · [GitHub](https://github.com/0xsero) · [REAP paper](https://arxiv.org/abs/2510.13999) · [Cerebras REAP](https://huggingface.co/collections/cerebras/cerebras-reap)
+# DeepSeek-V4-Flash-162B
+REAP-pruned [deepseek-ai/DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash).
+## At a glance
+| | |
+|---|---|
+| Base model | [deepseek-ai/DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) |
+| Format | BF16 |
+| Total params | **162B** |
+| Active / token | — |
+| Experts / layer | 144 |
+| Layers | 43 |
+| Hidden size | 4096 |
+| Context | 1,048,576 |
+| On-disk size | 94 GB |
+## Which variant should I pick?
+| Variant | Format | Link |
+|---|---|---|
+| `DeepSeek-V4-Flash-162B` **(this)** | BF16 | [link](https://huggingface.co/0xSero/DeepSeek-V4-Flash-162B) |
+| `DeepSeek-V4-Flash-162B-GGUF` | GGUF | [link](https://huggingface.co/0xSero/DeepSeek-V4-Flash-162B-GGUF) |
+| `DeepSeek-V4-Flash-180B` | BF16 | [link](https://huggingface.co/0xSero/DeepSeek-V4-Flash-180B) |
+| `DeepSeek-V4-Flash-180B-GGUF` | GGUF | [link](https://huggingface.co/0xSero/DeepSeek-V4-Flash-180B-GGUF) |
+| `DeepSeek-V4-Flash-213B` | BF16 | [link](https://huggingface.co/0xSero/DeepSeek-V4-Flash-213B) |
 **162B parameters | K144 REAP-pruned | 200K context | no speculative decoding**
 ## License
 MIT for the serving recipe and tooling. The base model weights follow the DeepSeek V4 Flash license. Review it before use.
+## License & citation
+License inherited from the base model.
+```bibtex
+@misc{lasby2025reap,
+  title  = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
+  author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
+  year   = {2025}, eprint = {2510.13999}, archivePrefix = {arXiv}
+}
+```
+## Sponsors
+Made possible by **NVIDIA · TNG Technology · Lambda · Prime Intellect · Hot Aisle**.