Model

Step 3.7 Flash

OPEN

StepFun · Step family · released May 29, 2026

Newest StepFun release (successor to Step 3.5 Flash), adding native multimodal (image/GUI/document) support that the text-only Step 3.5 Flash lacked. Not previously in corpus (corpus's newest StepFun entry is Step-3 from 2025-07-31). No shipped 'Step 4' flagship found -- only reported as in training.

ReasoningCodingVisionFunction callingTool useAgentic

2212.7

Elo · rank #61

Parameters

198B

Active params

11B (MoE)

Context

262K tokens

Architecture

Sparse Mixture-of-Experts with 1.8B vision encoder + 196B language backbone, 198B total / 11B active

License

Apache 2.0

Languages

—

API price (in/out)

$0.2 / $1.15

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath95.0%#19

best: GPT-5.2 · 100.0%

BrowseCompAgents75.8%#22

best: Kimi K3 · 91.2%

GDPval-AAAgents1415.8#28

best: Claude Fable 5 · 1932

GPQA DiamondReasoning78.4%#95

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning49.7%#10

best: Claude Opus 5 · 64.7%

MMMU-ProVision75.0%#29

best: Claude Opus 4.7 · 85.5%

SWE-bench ProCoding56.3%#29

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding76.5%#37

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding59.6%#44

best: Gemini 3.8 Flash · 89.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA134.6 GB1× H200 141GB

LoRA421.7 GB4× H200 141GB

Full fine-tune3177.9 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $5.69 (1× H200 141GB)

Plan a fine-tune →

Step family

Elo progression across releases

Step 3.7 FlashMay 20262212.7

Resources

HFstepfun-ai/Step-3.7-Flash DOCShttps://github.com/stepfun-ai/Step-3.7-Flash APIAvailable via StepFun API

API price $0.2/$1.15 · each benchmark row carries its own source badge (see methodology)