LLM Reference

Step 3.7 Flash

Researched today

Last refreshed 2026-05-29. Next refresh: weekly.

Open SourceMultimodalCodingRAGAgentsLong contextVisionJSON / Tool use

Step 3.7 Flash is worth evaluating for coding, rag, and agents when its provider route and context window match the workload.

Decision context: Coding task fit, 3 tracked provider routes, and research from 2026-05-29.

Use it for

  • Teams evaluating coding, rag, and agents
  • Workloads that can use a 256k context window
  • Buyers comparing 3 tracked provider routes

Do not use it for

  • Workloads where another current model has stronger sourced task evidence

Cheapest output

$1.15

OpenRouter per 1M tokens

Provider routes

3

Tracked API hosts

Quality / dollar

Grade C

Ranked by benchmark score divided by cheapest output price

Freshness

2026-05-29

Researched today

fresh

Top use-case fit

Coding

Q/$ C

1 relevant benchmark in the decision map.

RAG

Included by capability and metadata signals in the decision map.

Agents

Included by capability and metadata signals in the decision map.

Provider price ladder

Compare all 3
ProviderInput / 1MOutput / 1MCacheRoute
OpenRouter$0.200$1.15-
Serverless
StepFun$0.200$1.15read $0.040
Serverless
NVIDIA NIM---
ProvisionedPartial

Benchmark peer barsfor Coding

Migration checks

No linked migration route is available for this model yet.

About

Step 3.7 Flash is StepFun's open-weights multimodal Mixture-of-Experts model for agentic coding, tool use, long-context reasoning, image understanding, and video understanding. It combines a 196B-parameter language backbone with a 1.8B-parameter vision encoder, activates about 11B parameters per token, supports a 256K-token context window, and exposes low, medium, and high reasoning levels for speed/depth tradeoffs. StepFun reports leading open-model results on ClawEval-1.1, SimpleVQA with Search, and SWE-bench Pro at launch. Weights are available on Hugging Face under Apache 2.0.

Step 3.7 Flash has a 256k-token context window.

Step 3.7 Flash input tokens at $0.2/1M, output at $1.15/1M.

Capabilities

VisionMultimodalReasoningFunction CallingTool UseStructured OutputsPrompt Caching

Benchmark Scores(14)

Scores are benchmark-specific and are direction-aware: the same numeric gap can mean very different outcomes across suites. Use the leaderboard context and this model's provider route to decide whether the winning margin is meaningful for your workload.
BenchmarkScoreVersionSource
ClawEval-1.167.11st among open models at releasehttps://static.stepfun.com/blog/step-3.7-flash/
SimpleVQA with Search Tool79.21st at release (GPT-5.5: 79.1)https://static.stepfun.com/blog/step-3.7-flash/
V* with Python95.32nd at release (Kimi K2.6: 96.9)https://huggingface.co/stepfun-ai/Step-3.7-Flash
SWE-bench Pro56.32nd at release (Claude Opus 4.7: 64.3, GPT-5.5: 58.6)https://static.stepfun.com/blog/step-3.7-flash/
Terminal-Bench 2.159.5Comparison: Step 3.5 Flash 53.37%, DeepSeek V4 Flash 62.0%, Gemini 3.5 Flash 76.2%, GPT-5.5 82.7%, Claude Opus 4.7 69.4%https://static.stepfun.com/blog/step-3.7-flash/
Toolathlon49.5https://huggingface.co/stepfun-ai/Step-3.7-Flash
Humanity's Last Exam47.2https://huggingface.co/stepfun-ai/Step-3.7-Flash
GDPval-AA45.8https://huggingface.co/stepfun-ai/Step-3.7-Flash
WorldVQA58.1Comparison: Kimi K2.6 55.98%https://static.stepfun.com/blog/step-3.7-flash/
HR-Bench 4K89.1Comparison: Kimi K2.6 91.25%https://static.stepfun.com/blog/step-3.7-flash/
Android Daily61.9Comparison: Gemini 3 Flash 63.21%https://static.stepfun.com/blog/step-3.7-flash/
DeepSearchQA92.8https://static.stepfun.com/blog/step-3.7-flash/
BrowseComp75.8https://static.stepfun.com/blog/step-3.7-flash/
ResearchRubrics71.7https://static.stepfun.com/blog/step-3.7-flash/

API Versions

step-3.7-flash

Rankings

Specifications

FamilyStep
Released2026-05-29
Parameters198B (11B active)
Context256k
ArchitectureMixture of Experts
Specializationgeneral
LicenseApache 2.0
Trainingpretrained

Created by

One of China's leading AI 'Six Tigers'.

Shanghai, China
Founded 2023
Website