DEV.co
Open-Source LLM · stepfun-ai

Step-3.5-Flash

Step 3.5 Flash is a 196B parameter sparse Mixture-of-Experts (MoE) foundation model from StepFun that activates ~11B parameters per token. It targets reasoning, coding, and agentic workloads with a 256K context window, achieving 100–350 tokens/second throughput via Multi-Token Prediction. The model is Apache 2.0 licensed, ungated, and available via OpenRouter and StepFun's API, as well as downloadable for local deployment.

Source: HuggingFace — huggingface.co/stepfun-ai/Step-3.5-Flash
199.4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
175.3k
Downloads (30d)

Key facts

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

FieldValue
Developerstepfun-ai
Parameters199.4B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads175.3k
Likes824
Last updated2026-03-17
Sourcestepfun-ai/Step-3.5-Flash

What Step-3.5-Flash is

Step 3.5 Flash is a 45-layer transformer with 4,096 hidden dimensions, 128,896 token vocabulary, and sparse MoE routing (288 experts per layer, Top-8 selection). It supports 256K context via hybrid 3:1 Sliding Window Attention, generates 4 tokens per forward pass (MTP-3), and is optimized for inference efficiency on consumer to enterprise hardware. Model card provides arxiv references (2602.10604, 2601.05593, 2507.19427) for architecture and evaluation methodology.

Quickstart

Run Step-3.5-Flash 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/Step-3.5-Flash")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

Agentic AI systems and tool-use workflows

Sparse MoE architecture enables fast reasoning loops. Benchmarks show 88.2% τ²-Bench, 84.5% GAIA (no file), and support for complex multi-step agent tasks with context management.

Software engineering and code generation

74.4% SWE-bench Verified and 86.4% LiveCodeBench-V6 demonstrate strong coding ability. Low latency (100–350 tok/s) makes it suitable for real-time IDE integration and code analysis.

Long-document reasoning and analysis

256K context window with efficient attention (3:1 SWA ratio) supports large codebases, research papers, and knowledge bases without prohibitive memory cost.

Running & fine-tuning it

ESTIMATE: ~40–60 GB VRAM for full-precision inference on a single GPU (11B active params per token); quantized (int8/int4) deployment likely requires 20–30 GB. Card mentions M4 Max and NVIDIA DGX Spark; precise memory footprint and multi-GPU requirements are not stated. Requires review of official deployment guides for accuracy.

Model card does not document LoRA, QLoRA, or instruction-tuning feasibility. MoE routing and multi-token prediction architecture may require specialized fine-tuning frameworks. Requires review of StepFun's fine-tuning documentation or community forks (GitHub noted in card).

When to avoid it — and what to weigh

  • Latency-critical applications requiring sub-100ms inference on resource-constrained devices — While fast relative to dense 196B models, absolute latency numbers are not provided. Deployment on edge/mobile devices is not mentioned as supported.
  • Applications requiring real-time multilingual support across low-resource languages — Vocabulary (128,896 tokens) and benchmarks focus on English and Chinese. Coverage for other languages is not documented.
  • Production deployments where model context-window stability is critical — Card notes 3:1 SWA hybrid approach for long context; full evaluation of context window boundaries and potential degradation at limits is not detailed.
  • Regulatory environments requiring audited model training data provenance or bias mitigation documentation — Model card does not disclose training data composition, filtering, or formal bias evaluation results.

License & commercial use

Apache License 2.0 (permissive OSI license). Permits commercial use, modification, and distribution under Apache 2.0 terms (attribution and NOTICE file required). No additional proprietary restrictions mentioned in card.

Apache 2.0 is a standard permissive OSI license explicitly allowing commercial use. Model is ungated. However, API access via OpenRouter or StepFun platform entails separate commercial terms (verify with each provider). Self-hosted deployment for commercial purposes is permitted under Apache 2.0 provided license terms are retained.

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

Card does not provide security audit, adversarial robustness evaluation, or documented mitigations for jailbreaking or prompt injection. MoE architecture may present novel attack surfaces (expert routing). Local deployment claim emphasizes data privacy but does not detail model safety measures. Recommend reviewing third-party red-teaming reports and conducting internal security evaluation before production deployment.

Alternatives to consider

DeepSeek V3.2

Similar MoE sparse activation (37B active from 671B) with comparable reasoning benchmarks (AIME 93.1 vs. Step 97.3). Larger expert pool and slower generation (33 tok/s). Consider if maximum reasoning depth outweighs latency.

GLM-4.7

Dense 32B activation, similar generation speed (100 tok/s) and context support (256K). Simpler to serve and fine-tune; lower reasoning on AIME (95.7 vs. 97.3) and SWE-bench (73.8 vs. 74.4). Consider if deployment simplicity is valued over sparse efficiency.

MiniMax M2.1

Smaller MoE footprint (10B active from 230B), comparable speed (100 tok/s). Lower reasoning benchmarks (AIME 83.0, HMMT 2025 Feb 71.0) but lower compute cost. Consider for budget-constrained deployments prioritizing speed over reasoning depth.

Software development agency

Ship Step-3.5-Flash with senior software developers

Assess your hardware requirements, evaluate local vs. API deployment, and pilot with OpenRouter's free tier. Contact our AI engineering team to design agentic workflows and optimize inference for your workload.

Talk to DEV.co

Related open-source tools

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

Step-3.5-Flash FAQ

Can I use Step 3.5 Flash commercially in a self-hosted setup?
Yes, Apache 2.0 permits commercial use, modification, and redistribution provided you retain the Apache 2.0 license and NOTICE file. If using the OpenRouter or StepFun API, review their commercial terms separately.
What GPU do I need to run Step 3.5 Flash locally?
Card mentions M4 Max and NVIDIA DGX Spark as reference hardware but does not specify minimum VRAM. Estimate 40–60 GB for full precision; quantized versions (int8/int4) likely 20–30 GB. Exact requirements depend on batch size and inference framework. Requires verification with official deployment documentation.
Does Step 3.5 Flash support fine-tuning or instruction-tuning?
Not clearly stated in the model card. MoE architecture may require specialized frameworks (e.g., vLLM or custom RL setup noted for agentic training). Consult StepFun GitHub repo or official documentation for LoRA/QLoRA support.
How does Step 3.5 Flash compare to GPT-4 or Claude 3.5 in real-world agent tasks?
Benchmarks show competitive scores (GAIA 84.5%, xbench-DeepSearch 83.7%, τ²-Bench 88.2%) on structured evaluations. Real-world performance depends on task specificity, context quality, and system prompt design. Proprietary models are not directly benchmarked in the provided table; internal testing recommended for your use case.

Software developers & web developers for hire

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 Step-3.5-Flash is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy Step 3.5 Flash?

Assess your hardware requirements, evaluate local vs. API deployment, and pilot with OpenRouter's free tier. Contact our AI engineering team to design agentic workflows and optimize inference for your workload.