Instructions to use jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6 with Transformers:

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

pipe = pipeline("text-generation", model="jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6

SGLang

How to use jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6 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 "jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6" \
    --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": "jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6",
		"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 "jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6" \
        --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": "jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6 with Docker Model Runner:
```
docker model run hf.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6
```

jhyuckkim commited on Aug 31

Commit

5522fb9

verified ·

1 Parent(s): 815bf46

Align naming and results with the published arXiv version

Browse files

Files changed (1) hide show

README.md +12 -12

README.md CHANGED Viewed

@@ -17,7 +17,7 @@ tags:
 > **Research artifact, not a general-purpose model.**
 > These weights are released so that the experiments in the paper *Pruning and
 > Distilling Mixture-of-Experts into Dense Language Models* can be reproduced.
-> Each student is distilled on a fixed, small token budget (3.8B tokens for
 > Qwen3, 0.3B tokens for DeepSeek-V2-Lite and GPT-OSS) purely so that expert
 > scoring and grouping methods can be compared under an equal budget. Absolute
 > quality is therefore far below the teacher and below pretrained models of the
@@ -35,20 +35,20 @@ https://github.com/krafton-ai/moe-to-dense
 ## Results
-Full comparison group for this architecture, DeepSeek-V2-Lite to dense, 0.3B-token distillation (paper Table 6). This model's row is in
 bold, and rows whose weights are also released link to them.
 | Configuration | Wino | Hella | ARC-E | ARC-C | MMLU | Avg |
 |---|---|---|---|---|---|---|
-| [DO-AWP, K=6](https://huggingface.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-DO-AWP-K6) | 60.3 | 41.0 | 53.7 | 28.2 | 28.7 | 42.39 |
-| SF, K=12 | 56.9 | 41.1 | 54.8 | 28.4 | 24.6 | 41.16 |
-| DO-AWP, K=12 | 59.0 | 41.5 | 51.7 | 26.2 | 26.9 | 41.07 |
-| AWP, K=12 | 57.1 | 40.9 | 52.9 | 27.4 | 26.4 | 40.93 |
-| CP, K=12 | 55.6 | 40.2 | 53.5 | 26.6 | 26.8 | 40.53 |
-| [AWP, K=6](https://huggingface.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-AWP-K6) | 56.8 | 38.6 | 51.0 | 27.5 | 28.1 | 40.37 |
 | [SF, K=6](https://huggingface.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-SF-K6) | 54.9 | 38.9 | 52.4 | 27.1 | 26.9 | 40.04 |
 | **CP, K=6** | **53.0** | **36.9** | **49.4** | **25.7** | **25.3** | **38.07** |
-| Random MLP + teacher attention | 50.6 | 25.6 | 30.6 | 20.9 | 23.6 | 30.25 |
 | Random initialization | 50.1 | 25.4 | 28.8 | 24.1 | 22.9 | 30.27 |
 | Teacher (DeepSeek-V2-Lite) | 76.2 | 80.5 | 84.4 | 56.3 | 58.0 | 71.09 |
@@ -63,8 +63,8 @@ Downstream accuracy is Winogrande 5-shot, HellaSwag 10-shot, ARC-Easy 25-shot, A
 | Distillation data | FineWeb-Edu (`sample-10BT`), 0.3B tokens |
 | Expert scoring | CP |
 | Experts kept (K) | 6 |
-| Grouping | pure pruning (K equals top-k, one expert per group) |
-| Down-projection scaling | uniform |
 ## Usage
@@ -78,7 +78,7 @@ tok = AutoTokenizer.from_pretrained("jhyuckkim/DeepSeek-V2-Lite-Dense-3B-CP-K6",
 ```
 `trust_remote_code=True` is required. The student keeps DeepSeek's own model
-class so that MLA attention and the shared expert are preserved unchanged, and
 `configuration_deepseek.py` / `modeling_deepseek.py` ship with the repo.
 ## Citation

 > **Research artifact, not a general-purpose model.**
 > These weights are released so that the experiments in the paper *Pruning and
 > Distilling Mixture-of-Experts into Dense Language Models* can be reproduced.
+> Each student is distilled on a fixed, small token budget (~4B tokens for
 > Qwen3, 0.3B tokens for DeepSeek-V2-Lite and GPT-OSS) purely so that expert
 > scoring and grouping methods can be compared under an equal budget. Absolute
 > quality is therefore far below the teacher and below pretrained models of the
 ## Results
+Full comparison group for this architecture, DeepSeek-V2-Lite to dense, 0.3B-token distillation (paper Table 17). This model's row is in
 bold, and rows whose weights are also released link to them.
 | Configuration | Wino | Hella | ARC-E | ARC-C | MMLU | Avg |
 |---|---|---|---|---|---|---|
 | [SF, K=6](https://huggingface.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-SF-K6) | 54.9 | 38.9 | 52.4 | 27.1 | 26.9 | 40.04 |
+| SF, K=12 | 56.9 | 41.1 | 54.8 | 28.4 | 24.6 | 41.16 |
 | **CP, K=6** | **53.0** | **36.9** | **49.4** | **25.7** | **25.3** | **38.07** |
+| CP, K=12 | 55.6 | 40.2 | 53.5 | 26.6 | 26.8 | 40.53 |
+| [ACP, K=6](https://huggingface.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-ACP-K6) | 56.8 | 38.6 | 51.0 | 27.5 | 28.1 | 40.37 |
+| ACP, K=12 | 57.1 | 40.9 | 52.9 | 27.4 | 26.4 | 40.93 |
+| [DO-ACP, K=6](https://huggingface.co/jhyuckkim/DeepSeek-V2-Lite-Dense-3B-DO-ACP-K6) | 60.3 | 41.0 | 53.7 | 28.2 | 28.7 | 42.39 |
+| DO-ACP, K=12 | 59.0 | 41.5 | 51.7 | 26.2 | 26.9 | 41.07 |
+| Random FFN + teacher attn | 50.6 | 25.6 | 30.6 | 20.9 | 23.6 | 30.25 |
 | Random initialization | 50.1 | 25.4 | 28.8 | 24.1 | 22.9 | 30.27 |
 | Teacher (DeepSeek-V2-Lite) | 76.2 | 80.5 | 84.4 | 56.3 | 58.0 | 71.09 |
 | Distillation data | FineWeb-Edu (`sample-10BT`), 0.3B tokens |
 | Expert scoring | CP |
 | Experts kept (K) | 6 |
+| Grouping | pure pruning on the routed experts, 8 groups total (2 shared + 6 routed) |
+| Down-projection scaling | shared experts copied unscaled, routed groups scaled by average conditional probability |
 ## Usage
 ```
 `trust_remote_code=True` is required. The student keeps DeepSeek's own model
+class so that MLA attention and the shared experts are preserved unchanged, and
 `configuration_deepseek.py` / `modeling_deepseek.py` ship with the repo.
 ## Citation