Model

Step-3

OPEN

StepFun · Step-3 family · released Jul 31, 2025

Open-source 321B MoE (38B active) multimodal reasoning VLM, Apache-2.0, co-designed with MFA/AFD for cost-effective decoding.

ReasoningCodingVisionFunction callingTool useAgentic

1628.1

Elo · rank #97

Parameters

321B

Active params

38B (MoE)

Context

64K tokens

Architecture

MoE — 48 experts (3 routed + 1 shared per token), Multi-Matrix Factorization Attention (MFA) + Attention-FFN Disaggregation (AFD)

License

Apache-2.0

Languages

—

API price (in/out)

$0.57 / $1.42

Modalities

text · vision

Benchmark results

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

AIMEMath82.9%#69

best: GPT-5.2 · 100.0%

GPQA DiamondReasoning73.0%#112

best: GPT-5.6 · 94.6%

LiveCodeBenchCoding67.1%#78

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVisionVision64.8%#13

best: Seed 2.1 Pro · 92.6%

MMMUVision74.2%#39

best: Claude Fable 5 · 89.3%

Run it locally

VRAM @ Q4

177 GB

VRAM @ FP16

642 GB

Fits on (Q4)

M3 Ultra 512GBB200 192GB

321B total params must stay resident even with 38B active — infeasible on a single consumer GPU; needs multi-GPU/datacenter serving.

Quantizations

FP8

Step-3 family

Elo progression across releases

Step-3Jul 20251628.1

Resources

HFstepfun-ai/step3 GHstepfun-ai/Step3 DOCSstepfun.ai/research/en/step3 APIAvailable via StepFun API

API price $0.57/$1.42 · each benchmark row carries its own source badge (see methodology)