Instructions to use jackmaxwil/DeepSeek-V4.1-Flash-REAP-75 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jackmaxwil/DeepSeek-V4.1-Flash-REAP-75 with Transformers:

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

pipe = pipeline("text-generation", model="jackmaxwil/DeepSeek-V4.1-Flash-REAP-75")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("jackmaxwil/DeepSeek-V4.1-Flash-REAP-75", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jackmaxwil/DeepSeek-V4.1-Flash-REAP-75 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jackmaxwil/DeepSeek-V4.1-Flash-REAP-75"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jackmaxwil/DeepSeek-V4.1-Flash-REAP-75",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/jackmaxwil/DeepSeek-V4.1-Flash-REAP-75

SGLang

How to use jackmaxwil/DeepSeek-V4.1-Flash-REAP-75 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 "jackmaxwil/DeepSeek-V4.1-Flash-REAP-75" \
    --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": "jackmaxwil/DeepSeek-V4.1-Flash-REAP-75",
		"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 "jackmaxwil/DeepSeek-V4.1-Flash-REAP-75" \
        --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": "jackmaxwil/DeepSeek-V4.1-Flash-REAP-75",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use jackmaxwil/DeepSeek-V4.1-Flash-REAP-75 with Docker Model Runner:
```
docker model run hf.co/jackmaxwil/DeepSeek-V4.1-Flash-REAP-75
```

DeepSeek-V4.1-Flash-REAP-75

DeepSeek-V4.1-Flash with 75% of its routed experts removed by REAP (Router-weighted Expert Activation Pruning, arXiv:2510.13999). Each MoE layer keeps 96 of its 384 routed experts; the shared expert, attention, embeddings and every other weight are unchanged, and no weights were retrained after pruning.

This is an independent derivative of deepseek-ai/DeepSeek-V4.1-Flash-1017 (revision 1bae0e108899), made by jackmaxwil. It is not affiliated with or endorsed by DeepSeek. The base model is MIT-licensed and so is this checkpoint.

Sibling release: jackmaxwil/DeepSeek-V4.1-Flash-REAP-50. Calibration statistics, survivor lists and the eval data: jackmaxwil/DeepSeek-V4.1-Flash-REAP-stats.

What changed

n_routed_experts is 96 in config.json and inference/config.json; num_experts_per_tok stays 6.
The router weight and bias of every MoE layer are sliced to the surviving experts, and the surviving experts are renumbered 0..95 in their original order.
Expert weights keep the base model's FP4 format. keep-prune.json records the source index and survivor-list hashes; SHA256SUMS covers every file.

Run it exactly as the base model (see inference/); only the expert count differs.

How the experts were chosen

REAP scores each expert by its router gate weight times the norm of its output, averaged over the tokens routed to it, and keeps the top 96 per layer. Calibration ran on 2,946,031 tokens (24 sessions of up to 16,384-token records) of English and code, weighted toward agentic coding, terminal and tool use. Sources: nvidia/Open-SWE-Traces, nebius/SWE-rebench-V2-PRs, nvidia/Nemotron-Terminal-Corpus, Agent-Ark/Toucan-1.5M, CIRCL/vulnerability-scores, common-pile/stackexchange_filtered, HuggingFaceFW/finepdfs, corbt/enron-emails, allenai/WildChat-4.8M, HuggingFaceFW/finewiki. No Chinese text was used.

Evaluation

A held-out eval split from the same sources (108 sessions, 437,350 scored positions) was run through the full model and through the full model with the pruned experts masked out of the router. The table compares the pruned model with the full model as teacher: KL over the teacher's top 2,048 logits, top-1 agreement, and the change in NLL of the reference next token (negative means the pruned model gave the reference token higher probability).

Domain	Scored positions	Top-1 agreement	KL(teacher ‖ pruned)	NLL change (nats)
all	437,350	0.720	1.349	-0.131
agentic-python	63,310	0.635	2.055	-0.799
agentic-rust	49,149	0.625	2.055	-0.640
bash	50,094	0.882	0.447	-0.140
business	32,823	0.693	0.747	+0.535
business-email	15,382	0.696	1.011	+0.602
code-diffs	64,558	0.889	0.305	+0.232
cyber	30,806	0.658	1.615	+1.532
english-chat	15,439	0.632	0.962	+0.460
english-prose	18,873	0.757	0.968	+0.719
networking	32,674	0.820	0.411	+0.319
tool-use	64,242	0.587	2.807	-1.378

All four candidates scored in the same run:

Variant	Experts kept	Top-1 agreement	KL	NLL change
REAP-33 (not published)	256 / 384	0.842	0.511	+0.025
REAP-50	192 / 384	0.803	0.810	-0.138
REAP-67 (not published)	128 / 384	0.755	1.114	-0.109
REAP-75	96 / 384	0.720	1.349	-0.131

Read these numbers with care:

They come from router masking inside the full model. This exported checkpoint was then run on its own (on CPU, through the reference code with dequantize-then-matmul stand-ins for the FP8/FP4 kernels) over 5 of the eval sessions, 1,329 positions, and scored against the same teacher logits. It matched the masked-router numbers for those positions: KL 1.378 vs 1.385, NLL change -0.244 vs -0.225, top-1 agreement 0.676 vs 0.672. That check is teacher-forced scoring; free-running generation has not been tested yet.
The negative NLL changes on agentic and tool-use text mean the pruned model fits the reference traces (written by other models) a little better while moving away from the full model's own distribution. They are not evidence that pruning improves the model.
Only text was calibrated and evaluated. The vision path is untouched and untested.
There is no post-pruning fine-tuning. Expect more degradation on domains outside the calibration mix, and on Chinese in particular.

Downloads last month: 206

Safetensors

Model size

355B params

Tensor type

BF16

F32

F8_E4M3

Model tree for jackmaxwil/DeepSeek-V4.1-Flash-REAP-75

Base model

deepseek-ai/DeepSeek-V4.1-Flash-1017

Quantized

(203)

this model

Paper for jackmaxwil/DeepSeek-V4.1-Flash-REAP-75

REAP the Experts: Why Pruning Prevails for One-Shot MoE compression

Paper • 2510.13999 • Published Oct 15, 2025 • 24