Model

Step-4

OPEN

StepFun · Step family · released Sep 23, 2026

StepFun's September 23, 2026 flagship — the first Step-4 release and a return to open-weights flagships after the May Step-3.7-Flash mid-tier (610B/38B MoE, Apache-2.0). The launch table is StepFun's broadest yet: every shared row lands 4-8 points over 3.7-Flash, with a first-time OSWorld 3.0 filing (71.4) alongside ByteDance on the new revision. Capability flags keep the Step convention — agentic stays false (no first-party agent harness), and the $0.40/$1.60 API price undercuts the open-weights cohort.

ReasoningCodingVisionFunction callingTool useAgentic

2461.7

Elo · rank #38

Parameters

610B

Active params

38B (MoE)

Context

512K tokens

Architecture

610B-total/38B-active sparse MoE, Apache-2.0 open weights; 512K-token context

License

Apache-2.0

Languages

—

API price (in/out)

$0.4 / $1.6

Modalities

text · vision

Benchmark results

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

AIMEMath96.2%#10

best: GPT-5.2 · 100.0%

BrowseCompAgents82.6%#23

best: Kimi K3.1 · 92.4%

GDPval-AAAgents1489#29

best: Claude Fable 5 · 1932

GPQA DiamondReasoning84.6%#73

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning55.2%#7

best: Claude Opus 5 · 64.7%

MCP AtlasAgents79.8%#3

best: Kimi K3.1 · 85.6%

MMMU-ProVision81.3%#8

best: Claude Opus 4.7 · 85.5%

OSWorld 3.0Agents71.4%#2

best: Seed 2.2 Pro · 74.6%

SWE-bench ProCoding61.8%#16

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding81.4%#13

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding68.4%#37

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

QLoRA414.8 GB4× H200 141GB

LoRA1299.3 GB8× B200 192GB

Full fine-tune9790.5 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $19.65 (4× H200 141GB)

Plan a fine-tune →

Step family

Elo progression across releases

Step 3.7 FlashMay 20262177.0 Step-4Sep 20262461.7

Resources

HFstepfun-ai/step-4 GHstepfun-ai/Step4 DOCShttps://platform.stepfun.com/docs/step-4 APIAvailable via StepFun API

API price $0.4/$1.6 · each benchmark row carries its own source badge (see methodology)