Instructions to use stepfun-ai/Step-3.5-Flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use stepfun-ai/Step-3.5-Flash with Transformers:

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

pipe = pipeline("text-generation", model="stepfun-ai/Step-3.5-Flash", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("stepfun-ai/Step-3.5-Flash", trust_remote_code=True, dtype="auto")

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use stepfun-ai/Step-3.5-Flash with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "stepfun-ai/Step-3.5-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": "stepfun-ai/Step-3.5-Flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/stepfun-ai/Step-3.5-Flash

SGLang

How to use stepfun-ai/Step-3.5-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 "stepfun-ai/Step-3.5-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": "stepfun-ai/Step-3.5-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 "stepfun-ai/Step-3.5-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": "stepfun-ai/Step-3.5-Flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use stepfun-ai/Step-3.5-Flash with Docker Model Runner:
```
docker model run hf.co/stepfun-ai/Step-3.5-Flash
```

This is the most underreported model when it comes to agentic coding and intelligence

#35

by mayankiit04 - opened Mar 25

Discussion

mayankiit04

Mar 25

I am running a reap version of the same and still step 3.5 flash does really excel even compared to qwen series and minimax reap for my agentic coding in rust. Hope this team continues its good work., best of luck!!

deleted

Apr 8

Minimax (non-reaped) seems better to me, even at IQ2. M2.7 will wipe the floor with Step-3.5-Flash, so looking forward to Step-3.6-Flash.

Otherwise, absolutely; probably still better than Qwen3.5 122B, for example. Though a large part of the problem was that inference engines like llama.cpp and vllm couldn't run them for a long time, I think. Arguably that's the same issue: they were underappreciated.

BingoBird

25 days ago

https://x0.at/vURT.webp comparison to SOTA models. Very very strong performance!

deleted

25 days ago

•

edited 25 days ago

https://x0.at/vURT.webp comparison to SOTA models. Very very strong performance!

Not sure what that's measuring, but dollars isn't really a useful comparison for open weight models. Need the agentic intelligence per token, really. On https://artificialanalysis.ai , the "Agentic Index" and "Intelligence vs. Output Tokens" charts are useful:

https://artificialanalysis.ai/?intelligence=agentic-index&intelligence-efficiency=intelligence-efficiency-vs-output-tokens&models=gpt-oss-120b%2Cgpt-oss-20b%2Cgemma-4-31b%2Cmistral-small-4%2Cdeepseek-v3-2-reasoning%2Cminimax-m2-7%2Cnvidia-nemotron-3-super-120b-a12b%2Ckimi-k2-6%2Ck-exaone%2Ck2-think-v2%2Cglm-5-1%2Cqwen3-5-397b-a17b%2Cstep-3-5-flash-0202&model-filters=open-source#artificial-analysis-intelligence-index-by-open-weights-proprietary

https://artificialanalysis.ai/?intelligence=agentic-index&intelligence-efficiency=intelligence-efficiency-vs-output-tokens&models=gpt-oss-120b%2Cgpt-oss-20b%2Cgemma-4-31b%2Cmistral-small-4%2Cdeepseek-v3-2-reasoning%2Cminimax-m2-7%2Cnvidia-nemotron-3-super-120b-a12b%2Ckimi-k2-6%2Ck-exaone%2Ck2-think-v2%2Cglm-5-1%2Cqwen3-5-397b-a17b%2Cstep-3-5-flash-0202&model-filters=open-source#intelligence-vs-output-tokens-used-in-artificial-analysis-intelligence-index

As you can see there, Step-3.5-Flash WAS good for an older model, but needs an update. MiniMax M2.7 currently outclasses it. MiniMax and StepFun have been collaborating though, and that's part of the reason for M2.7 being ahead, so we'll see what 3.6 is like.

One more thing to add is that StepFun are targeting 128GB machines with Flash: 128GB VRam (4x32GB GPUs, or 1xMac with 128GB unified RAM), so that's achievable for a lot of the huggingface/local llama crowd, and with a more aggressive quant, it's in reach for 96GB folks too. MiniMax will run in 96GB, but it's a tighter squeeze and a very aggressive quant.

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

Tap or paste here to upload images

· Sign up or log in to comment