Bonus: Top up now and we'll double your first deposit — get x2 credits instantly.
Get Started
Sign Up
Sign Up

Step 3.5 Flash: Self-Hosting & Deployment Guide

Step 3.5 Flash is StepFun’s sparse Mixture-of-Experts language model that delivers frontier-level reasoning and agentic capabilities while remaining highly efficient and fast for production use.

Browse available models

What is Step 3.5 Flash?

Step 3.5 Flash is a sparse Mixture-of-Experts large language model from StepFun designed to combine frontier-level reasoning with high-throughput, low-latency inference. It is mainly used for complex reasoning tasks, code generation, and agentic workflows that benefit from its long context window and efficient token usage. The model is also applied to natural language processing, data analysis, and long-document or codebase understanding where cost and speed are critical. It belongs to StepFun’s Step family of models and builds on the Step 3.x architecture and research line.

How to run Step 3.5 Flash yourself

Open weights (apache-2.0), 199.4B parameters. Below: how much memory the weights need at each quantization, which hardware fits them, and the commands to serve it.

Step 3.5 Flash VRAM requirements

Precision / quantizationWeightsFigureSource
BF16 (full precision)398.8 GBEstimate199.4B params × 2 bytes
FP8208.6 GBFile sizestepfun-ai/Step-3.5-Flash-FP8
AWQ 4-bit126.4 GBFile sizenp-deploys/Step-3.5-Flash-AWQ-4bit
MLX 4-bit110.8 GBFile sizemlx-community/Step-3.5-Flash-4bit

File size = actual download size on Hugging Face. Estimate = parameter count × bytes per weight. Weights only — the KV cache for long contexts needs extra memory; the fit tables below add 15% headroom as a rule of thumb.

GPU requirements: deploy Step 3.5 Flash on your hardware

HardwareBF16 (full precision)FP8AWQ 4-bitMLX 4-bit
NVIDIA H200141 GB4×2×2×1×
NVIDIA H10080 GB6×3×2×2×
NVIDIA A100 80GB80 GB6×3×2×2×
NVIDIA L40S48 GB—5×4×3×

GPUs needed per variant.

HardwareMLX 4-bit
Mac, 512 GBM3 UltraFits
Mac, 192 GBM2 UltraFits
Mac, 128 GBM4 Max / M3 Max—
Mac, 64 GBM4 Pro / Max—

Assumes ~75% of unified memory is usable by the GPU.

HardwareBF16 (full precision)FP8AWQ 4-bitMLX 4-bit
AWS p5.48xlarge8× H100 · 640 GBFitsFitsFitsFits
AWS p4de.24xlarge8× A100 80GB · 640 GBFitsFitsFitsFits
AWS g6e.12xlarge4× L40S · 192 GB——FitsFits
Google Cloud a3-highgpu-8g8× H100 · 640 GBFitsFitsFitsFits
Azure ND H100 v58× H100 · 640 GBFitsFitsFitsFits

Whole-instance GPU memory; public instance specs.

Serve it yourself

pip install -U vllm
vllm serve stepfun-ai/Step-3.5-Flash
pip install 'sglang[all]'
python -m sglang.launch_server --model-path stepfun-ai/Step-3.5-Flash
pip install -U mlx-lm
mlx_lm.generate --model mlx-community/Step-3.5-Flash-4bit --prompt "Hello"

Commands use the official repositories above; vLLM and SGLang expose an OpenAI-compatible endpoint.

Not on LLM.API yet

We do not route this model through our API at the moment. Browse the models you can call today — most workloads have a close match already live.

Browse available models

Providers

Different companies host the same model. OpenRouter routes your request to one of them based on the routing mode you pick — Balanced (price + speed), Nitro (fastest), or Exacto (highest tool-calling accuracy).

List price by provider

InputOutput
SiliconFlow$0.0001 in
$0.0003 out

Provider list prices; the LLM.API discount applies on top.

Provider Input Output Cache read /M Latency Throughput Uptime
SiliconFlowin $0.0001; out $0.0003 per 1M tokens262K tokenstools, streaming, reasoning

List prices from third-party hosts in the OpenRouter public catalogue (stepfun/step-3.5-flash); LLM API does not serve this model and applies no discount. Updated weekly. Last updated 4 Oct 2026.

Try this model

Test Step 3.5 Flash right here — free to start.

Step 3.5 Flash
Hi! Want to test the model?

Suggestions for your first prompt

Not available on LLM API yet

We do not route this model through our API at the moment, so there is no endpoint or code snippet for it yet. Browse the models you can call today — most workloads have a close match already live.

Browse available models

Step 3.5 Flash: Self-Hosting & Deployment Guide in production

What the model is good at.

  • Conversational Chat

    Handles general-purpose dialogue, explanations, brainstorming, and question answering, optimized for fast, low-latency text generation and responses.

  • Structured JSON Output

    Generates well-formed JSON and structured text suitable for programmatic consumption, including configuration data, responses, and tool outputs.

  • Multilingual Translation

    Translates between multiple natural languages, leveraging its large context and reasoning capabilities to preserve meaning and style.

  • Code Generation

    Writes and edits code, explains snippets, and assists with debugging across common programming languages, tuned for agentic coding workflows.

  • Long-Context Reasoning

    Performs reasoning and synthesis over very long texts using its 256K-token context window for documents, logs, or multi-step analyses.

What you get when you call Step 3.5 Flash: Self-Hosting & Deployment Guide through LLM.API.

  • Intelligent AI Routing

    Automatically route each request to the optimal model based on latency, cost, and quality—without changing your integration or redeploying services.

    One endpoint, every model.

  • Cost-Aware Orchestration

    Control spend with dynamic model selection, price ceilings, and transparent usage metrics so you can scale AI features without runaway cloud bills.

    Max performance, minimal spend.

  • Resilient Fallback Logic

    Define automatic failover chains across providers so timeouts, rate limits, or outages don’t break your workloads or user experience.

    Stay online, even upstream.

  • End-to-End Observability

    Trace every call across providers with logs, metrics, and latency breakdowns to debug faster and optimize prompt, model, and routing decisions.

    See every token’s journey.

  • Task-Level Orchestration

    Describe high-level tasks once and let LLM.API handle tool calls, multi-step flows, and provider choices for consistent results across environments.

    Think tasks, not models.

  • High-Throughput Batch APIs

    Send thousands of operations in a single request with controlled concurrency, retries, and deduplication to power large-scale inference pipelines efficiently.

    Ship at batch scale.

When to Use — When NOT to Use

Use it if...

  • You need a low-cost general-purpose model for everyday chat and task automation.
  • You need fast responses for lightweight classification, tagging, and basic extraction tasks.
  • Your use case involves generating short marketing copy, social content, or product descriptions.
  • Your use case involves simple code help, like small snippets, comments, or refactors.
  • You need an inexpensive model to power AI features in consumer or internal tools.
  • You need a model to summarize short documents, tickets, or support conversations efficiently.
  • Your use case involves multilingual but simple Q&A where perfect nuance is not critical.

Avoid if...

  • You need frontier-level reasoning quality for complex problem solving or strategic decision support.
  • Your workload requires highly reliable long-context handling across very large documents or codebases.
  • You need state-of-the-art code generation for complex systems, architectures, or optimization-heavy tasks.
  • Your workload requires strong domain expertise in sensitive areas like medicine, law, or finance.
  • You need advanced tool use, multi-step planning, or orchestration across many external systems.
  • Your workload requires carefully controlled style-mimicry, safety-tuned outputs, and strict content controls.
  • You need highest possible model quality for user-facing flagship features or premium products.

COMMUNITY

What developers say about Step 3.5 Flash

LMArena#143 · Elo 1403Arena coding#119 · Elo 1436Speed—Sources: LMArena, Artificial Analysis

Summarised from public developer discussions and write-ups. Opinions are the sources’, not LLM.API’s.

  • Developers on Hacker News report the model is underrated, highlighting its high context efficiency on Mac hardware and its ability to outperform models like GLM-4.7 and MiniMax 2.5 in local 4-bit testing.
  • Developers on Hugging Face praise the model's speed and planning capabilities for agentic coding, noting it successfully created and executed multi-phase plans for complex projects better than previous local models.
  • Users on developer blogs and GitHub note that while the model handles structured code and multiple languages stably, it can occasionally experience bugs like infinite reasoning loops or tool-calling errors.

SOURCES

Frequently Asked Questions

  • What is Step 3.5 Flash?

    Step 3.5 Flash is a fast, cost-efficient StepFun language model for general-purpose text generation and reasoning, accessible through the LLM.API gateway.

  • What is Step 3.5 Flash best suited for?

    Step 3.5 Flash is best for high-throughput tasks like chatbots, agents, data processing, and lightweight reasoning where low latency and low cost matter.

  • What is the context window of Step 3.5 Flash?

    Step 3.5 Flash supports a context window of up to 32K tokens, including both prompt and completion tokens.

  • How fast is Step 3.5 Flash in terms of latency?

    Step 3.5 Flash is optimized for low latency, typically returning first tokens within a few hundred milliseconds depending on prompt size and load.

  • What modalities does Step 3.5 Flash support?

    Step 3.5 Flash is a text-only model that accepts textual prompts and returns textual completions.

  • How do I call Step 3.5 Flash through LLM.API?

    Specify the provider as "StepFun" and the model name "step-3.5-flash" in your LLM.API request, sending standard chat or completion payloads.

  • How is Step 3.5 Flash priced on LLM.API?

    On LLM.API, Step 3.5 Flash is billed per 1,000 tokens of prompt and completion; check the LLM.API pricing page for exact rates.

  • How does Step 3.5 Flash compare to larger StepFun models?

    Compared to larger StepFun models, Step 3.5 Flash is cheaper and faster but generally less capable on complex reasoning and intricate long-context tasks.

  • Does Step 3.5 Flash support streaming responses via LLM.API?

    Yes, Step 3.5 Flash can stream tokens incrementally when you enable streaming mode in your LLM.API request.

  • What are the main limitations of Step 3.5 Flash?

    Step 3.5 Flash may struggle with very long multi-step reasoning, domain-expert tasks, strict factual accuracy, and does not access external tools or the internet.

COMPARE

Competitive Models

Get one key to every model

Swap your API key. Keep your code.

Browse available models Talk to sales