DEV.co
Open-Source LLM · stepfun-ai

step3

Step3 is a 321B-parameter multimodal model (38B active via Mixture-of-Experts) from StepFun that processes both images and text. It uses custom attention mechanisms designed to reduce inference costs while maintaining strong reasoning performance. The model is open-source under Apache 2.0 and can be self-hosted or accessed via StepFun's API.

Source: HuggingFace — huggingface.co/stepfun-ai/step3
321B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
175.8k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerstepfun-ai
Parameters321B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / taskimage-text-to-text
Gated on HuggingFaceNo
Downloads175.8k
Likes166
Last updated2026-01-29
Sourcestepfun-ai/step3

What step3 is

Step3 is a vision-language model built on a 48-expert MoE architecture with 3 experts selected per token, employing Multi-Matrix Factorization Attention (MFA) and Attention-FFN Disaggregation (AFD) for efficiency. It supports 65,536 token context, uses bf16 precision for inference, and integrates the Deepseek V3 tokenizer. The model is available in safetensors format with vLLM and SGLang optimizations. Architecture: 61 layers (5 dense), 7168 hidden dimension, 64 query heads, 256 head dimension.

Quickstart

Run step3 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="stepfun-ai/step3")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Document and visual analysis at scale

65K context window and 38B active parameters enable processing lengthy documents with embedded images, OCR tasks, and multi-page reasoning. MoE routing reduces per-token cost.

Cost-sensitive multimodal inference services

StepFun's emphasis on decoding efficiency and hardware flexibility makes Step3 suitable for production APIs and on-device deployments where inference cost per request is critical.

Custom vision-language applications

Transformers integration with custom code support allows fine-tuning for domain-specific image-text tasks (medical imaging, industrial inspection, content moderation) on limited hardware.

Running & fine-tuning it

Estimated 130–160 GB GPU VRAM for bf16 inference (assuming ~0.5 bytes per parameter after MoE compression). Multi-GPU setups with device_map='auto' are supported. Requires torch>=2.1.0, transformers>=4.54.0, Python 3.10+. vLLM and SGLang recommended for optimized serving. No quantized variants (int8, int4, GGUF) are documented; bf16 and block-fp8 are the published formats.

No LoRA/QLoRA details in the model card. Standard transformers fine-tuning is possible but may be prohibitively expensive at 321B parameters. LoRA support and low-rank fine-tuning strategies are Unknown; users should consult the GitHub repository or contact StepFun directly. MoE routing may require specialized adaption techniques.

When to avoid it — and what to weigh

  • Real-time latency is paramount — MoE and custom attention patterns add routing and computational overhead. Step3 may not match single-model latency benchmarks; verify with vLLM/SGLang profiling against your hardware.
  • Minimal VRAM budget or mobile/edge deployment — 321B total parameters require significant VRAM even with 8-bit quantization. The model card does not mention GGUF/int4 support; bf16 is the stated format.
  • Proprietary fine-tuning constraints — If your deployment requires strict model IP isolation or cannot accept custom_code execution, Step3's reliance on trust_remote_code in transformers may conflict with your security posture.
  • Proven, battle-tested model maturity required — Step3 is recent (model card dated 2025). Production users should validate benchmarks independently and monitor for patches; no adoption or incident history is available.

License & commercial use

Apache License 2.0 (Apache-2.0). This is a permissive, OSI-approved license covering both code repository and model weights. Permits commercial use, modification, and redistribution under the terms of the license.

Apache 2.0 is a permissive OSI license that explicitly allows commercial use. No gating restrictions apply (gated=false). You may use Step3 in production, resell services, and modify the model, provided you include a copy of the Apache 2.0 license and any modifications to the code. No patent indemnity or enterprise support is implied by the license alone; contact StepFun for commercial support or indemnification SLAs.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Step3 requires trust_remote_code=True in transformers, meaning custom model code is executed. Verify the GitHub repository for code integrity and dependencies. No documented security audit, vulnerability disclosure policy, or hardening measures are stated. For production use, review the source on GitHub and consider code signing/supply-chain controls. Standard LLM risks apply: prompt injection, jailbreaking, and data leakage in multi-tenant environments require application-level safeguards.

Alternatives to consider

Llama 3.1 Vision (405B or smaller variants)

Strong multimodal reasoning, larger adoption base, and more transparent commercial license (Llama 3.1 Community License). Comparable or better benchmarks but higher inference cost without MoE optimization.

Qwen2-VL (72B)

Smaller footprint, proven production stability, and strong vision-language benchmarks. Better VRAM fit for cost-conscious deployments, though smaller context window (32K) and no MoE efficiency.

GPT-4 Vision API or Claude 3.5 Sonnet

If self-hosting is not required, managed APIs eliminate infrastructure burden, offer SLA support, and reduce operational risk. Trade-off: higher per-token cost and vendor lock-in.

Software development agency

Ship step3 with senior software developers

Step3 is Apache-licensed and ready for self-hosting or custom integration. Start with vLLM on an H100, evaluate benchmarks on your workload, and contact StepFun for production support or API access.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

step3 FAQ

Can I use Step3 commercially without restrictions?
Yes. Apache 2.0 is a permissive open-source license that permits commercial use, modification, and redistribution. You must include the license and disclose any code changes. No separate commercial agreement is required, but StepFun does not offer warranty or indemnification; contact them for enterprise SLAs if needed.
What GPU VRAM do I need to run Step3?
Approximately 130–160 GB for bf16 inference. An H100 (80 GB) or 2× A100s (80 GB each) with inter-GPU communication is practical. StepFun notes the model runs on both flagship and low-end accelerators; specific low-end configurations are not detailed. Use vLLM or SGLang for memory optimization.
Can I fine-tune Step3 on my own data?
Standard transformers fine-tuning is supported, but full-model training at 321B parameters is prohibitively expensive. LoRA/QLoRA feasibility is Unknown; consult the GitHub repository and StepFun's contact for guidance on parameter-efficient adaption strategies.
Is Step3 a recent model? How stable is it?
The model was released in 2025 (last modified 2026-01-29). It is recent and may not have production incident history like larger, established models. Verify benchmarks against your workload and monitor GitHub for patches. For mission-critical deployments, consider longer evaluation periods or fallback to more mature models.

Custom software development services

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If step3 is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy Step3?

Step3 is Apache-licensed and ready for self-hosting or custom integration. Start with vLLM on an H100, evaluate benchmarks on your workload, and contact StepFun for production support or API access.