Instructions to use deepseek-ai/DeepSeek-V4-Flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use deepseek-ai/DeepSeek-V4-Flash with Transformers:

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

pipe = pipeline("text-generation", model="deepseek-ai/DeepSeek-V4-Flash")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Flash")
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V4-Flash", device_map="auto")

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use deepseek-ai/DeepSeek-V4-Flash with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "deepseek-ai/DeepSeek-V4-Flash"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "deepseek-ai/DeepSeek-V4-Flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/deepseek-ai/DeepSeek-V4-Flash

SGLang

How to use deepseek-ai/DeepSeek-V4-Flash 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 "deepseek-ai/DeepSeek-V4-Flash" \
    --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": "deepseek-ai/DeepSeek-V4-Flash",
		"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 "deepseek-ai/DeepSeek-V4-Flash" \
        --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": "deepseek-ai/DeepSeek-V4-Flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use deepseek-ai/DeepSeek-V4-Flash with Docker Model Runner:
```
docker model run hf.co/deepseek-ai/DeepSeek-V4-Flash
```

Add LHTB eval result: 2/46 tasks solved (90-min budget, estimated)

#56

by zli12321 - opened about 13 hours ago

base: refs/heads/main

←

from: refs/pr/56

Discussion Files changed

-0

zli12321

about 13 hours ago

Adds the Long-Horizon Terminal-Bench (LHTB) result for DeepSeek V4 Flash to the LHTB benchmark leaderboard.

Score: 2 — tasks solved at reward >= 0.95, out of 46. The leaderboard metric is a task count, not a percentage.

Please read the caveat before merging, because this number is an estimate:


Agent	terminus-2 (official LHTB Harbor harness)
Run budget	3 hours
Leaderboard budget	90 minutes
Score at 90 min	2/46 (range 2-3, `riscv-core-debug` unresolved)
Score at full 3h	5/46

We only have a 3-hour run of DeepSeek V4 Flash on terminus-2, but every entry on this
leaderboard uses a 90-minute budget. Rather than publish the 3-hour number as if it were
comparable, the 90-minute figure is reconstructed from the run's interim verifier
checkpoints: LHTB re-verifies after each phase, so a task's reward at an earlier cutoff can
be read back. Ten tasks have no checkpoint that early and are counted as unsolved, which is
why 2 is a lower bound and 3 the upper.

This is not a substitute for actually running at 90 minutes -- the continuation prompt
tells the agent how much time remains, so a real shorter run would budget its effort
differently. Full artifacts and the per-budget curve (reward_curve.json) are published so
the derivation can be checked:
https://huggingface.co/datasets/IntelligenceLab/LHTB-leaderboard/tree/main/submissions/long-horizon-terminal-bench/1.0-extended/terminus-2__deepseek-v4-flash-3h

Submitted by the LHTB maintainers. Happy to withdraw this if you would rather wait for a
native 90-minute run, or to adjust the wording.

Add LHTB eval result: 2/46 tasks solved (90-min budget, estimated)125736e6

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

Ready to merge

This branch is ready to get merged automatically.

· Sign up or log in to comment