An Apache 2.0 open-weight Flash model for real-world agents. Step 3.7 Flash combines vision, coding, search, tool use, 256K context, ~11B active params, and up to 400 TPS.
Step 3.7 Flash is basically an efficiency-first model for real-world agents.
It’s fast (up to 400 TPS), handles vision + tool use quite well, and stays relatively stable on longer agent runs. Also plays nicely with common harnesses like @Claude Code, @OpenClaw and @Kilo Code.
@Stepfun's Step 3.7 Flash is one of the best open-weight models you can run right now.
Perfect timing to play with Kilo's new VS Code extension launched this month.
Report
stepfun is an interesting company to watch because they're operating in a space where the obvious question is always why this over Qwen or Mistral for open weight deployments. the vision plus tool use combination at this size is differentiated but the developer ecosystem and tooling support matters as much as the benchmarks at this point. what does the integration story actually look like, is it compatible with existing inference frameworks or does it need custom setup
Report
The efficiency first angle makes sense for agents. In real workflows, speed and stable tool use can matter more than raw model size. How does Step 3.7 Flash perform when a task needs multiple tool calls plus vision in the same run?
Replies
Flowtica Scribe
Hi everyone!
Step 3.7 Flash is basically an efficiency-first model for real-world agents.
It’s fast (up to 400 TPS), handles vision + tool use quite well, and stays relatively stable on longer agent runs. Also plays nicely with common harnesses like @Claude Code, @OpenClaw and @Kilo Code.
The weights are also open under Apache 2.0.
Love this kind of Flash model: not trying to be the biggest, but trying to be the one agents can actually run a lot.
Kilo Code
@Stepfun's Step 3.7 Flash is one of the best open-weight models you can run right now.
Perfect timing to play with Kilo's new VS Code extension launched this month.
stepfun is an interesting company to watch because they're operating in a space where the obvious question is always why this over Qwen or Mistral for open weight deployments. the vision plus tool use combination at this size is differentiated but the developer ecosystem and tooling support matters as much as the benchmarks at this point. what does the integration story actually look like, is it compatible with existing inference frameworks or does it need custom setup
The efficiency first angle makes sense for agents. In real workflows, speed and stable tool use can matter more than raw model size. How does Step 3.7 Flash perform when a task needs multiple tool calls plus vision in the same run?