Skip to main content
Back

StepFun's Step 3.7 Flash Runs 198B Vision Agents at One-Ninth Frontier Costs

StepFun's 198B Mixture-of-Experts model adds native vision, three reasoning gears, and an Apache 2.0 license that runs on a 128GB workstation.

Pro

Model, Full Access
  • StepFun released Step 3.7 Flash, a 198B MoE vision-language model, Apache 2.0, with ~11B active parameters per token.
  • Unsloth shipped Dynamic 2.0 GGUF quants, runnable on 128GB unified-memory machines at full 256K context.
  • Three reasoning gears (low, medium, high) selectable per API call via reasoning_effort.
  • Pricing: $0.20/M input and $1.15/M output, with Advisor Mode reportedly hitting ~$0.19 per agent task.
  • Scores 56.3% on SWE-Bench Pro and leads ClawEval-1.1 at 67.07% for tool-use reliability.
  • Available on StepFun API, OpenRouter, NVIDIA NIM, with vLLM, SGLang, and llama.cpp backends.

StepFun just dropped Step 3.7 Flash, an open-weight vision-language model aimed squarely at agent workloads, and Unsloth has already shipped a full set of GGUF quantizations so you can run it on a single high-memory workstation. The combination matters because it puts a frontier-tier agentic model into reach for teams that do not want to send every tool call to a closed API.

Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model that combines a 196B-parameter language backbone with a 1.8B-parameter vision encoder for native image understanding. Engineered for high-frequency production workloads, it activates approximately 11B parameters per token and delivers a throughput of up to 400 tokens per second. It supports a 256k context window and offers three selectable reasoning levels (low, medium, and high) so developers can easily balance speed, cost, and cognitive depth.

The sparse-MoE math behind the cost story

In a Mixture-of-Experts model, only a slice of the network fires for any given token. Only a subset of expert sub-networks fires per forward pass rather than the full network, which keeps inference compute closer to an 11B dense model while maintaining a 198B total parameter budget. That ratio is the entire pricing argument.

The model runs roughly 11B parameters of compute per token, prices output at $1.15 per million tokens, and per StepFun's own scaffold runs an agentic task for about $0.19 where a frontier model runs the same task for closer to $1.76. For teams running thousands of agent loops a month, that ratio is the headline. Input sits at $0.20 per million tokens with an 80% cache-hit discount.

Three reasoning gears, one API call

Rather than ship separate thinking and non-thinking checkpoints, StepFun exposes a single reasoning_effort parameter. The Chat Completions API uses reasoning_effort to control the effort level; the Messages API uses output_config.effort. The three settings map cleanly to task complexity:

  • Low for fast, cheap completions
  • Medium for balanced tasks
  • High