StepFun's Step 3.7 Flash, a 198-billion-parameter sparse MoE vision-language model, is now available in the Baseten Model Library in a hardware-efficient configuration (lower cost per token, smoother autoscaling).
We're excited to welcome StepFun to the Baseten Model Library! StepFun is doing incredible work building frontier-scale foundation models across text, vision, audio, and reasoning.
Step 3.7 Flash is StepFun’s debut model in our library, and we can't wait to see how teams use it.
Why Step 3.7 Flash matters
Step 3.7 Flash is StepFun's flagship multimodal reasoning model. Here's what makes it stand out:
Efficient architecture: A highly efficient, sparse MoE architecture (198B total parameters / 11B activated parameters). Of its 198 billion parameters, only 11 billion activate per token. Think of it as a large team where each request is handled by a small, specialized subset of experts, so most of the "brainpower" stays at rest.
Multimodal input: Native image and video input, meaning you won't need to juggle separate models for visual tasks.
Long context window: A generous 256k token context window to handle your larger inputs, such as large codebases.
Flexible reasoning: Versatile reasoning capabilities at low, medium, and high levels, suited for supporting agentic workflows like coding and tool calling.
What can you do with Step 3.7 Flash?
The unique combination of multimodal input, a long context window, and flexible reasoning unlocks a world of creative use cases. Here are a few that StepFun showcases:
Excited to try it for yourself? Getting started is easier than you might think with Baseten.
Lowering cost and simplifying deployment
FP8 quantization means the model fits in 198 GB, so we can serve it on 4×H100s instead of 8×H200s . This makes Step 3.7 Flash cheaper, easier to deploy, and still with plenty of headroom for production traffic.
The official vLLM recipe for deploying Step 3.7 Flash lists 8xH200s or 8xB200s as the hardware prerequisite. In the Baseten deployment from our Model Library, we're using 4×H100s, substantially less expensive per token and more readily available than the larger configs.
How did we know it was possible to use a smaller instance? The FP8-quantized version of Step 3.7 Flash shrinks each weight from 16 bits to 8 bits, cutting memory requirements roughly in half (learn more about FP8 quantization and how we quantize without sacrificing quality). As a result, the model weights alone require only 198 GB of memory. The 8×H200s and 8×B200s configs provide 1,128 GB and 1,536 GB of that memory respectively, far beyond what the quantized model needs.
But is ~122 GB of remaining memory enough to serve production traffic? Yes, for three reasons:
Mixture-of-experts (MoE) architecture: Activates only 11 billion parameters (out of 198 billion) per token. Think of it as a large team where each request is handled by a small, specialized subset of experts.
Hybrid attention: Keeps the memory footprint from growing much, even as conversations get longer (learn more here).
Speculative decoding: Predicts several tokens ahead in a single step and then verifies them in parallel (learn more here). Correct predictions are essentially "free," boosting throughput without quality loss.
Together, these mean 4×H100s not only fit the model, but have enough headroom for production workloads.
Get started with Step 3.7 Flash
Step 3.7 Flash is available in the Baseten Model Library as a Dedicated Inference deployment. Click here to deploy Step 3.7 Flash now!
Once deployed, generate an API Key, and follow StepFun's quickstart guide to start processing image, video, and text inputs.
