StepFun: Step 3.5 Flash

StepFun flash lane for quick multimodal reasoning and coding assistance

step-3.5-flash

AvailableStepFunChat

Input / Output: $0.10 / $0.30
Context: 256K
Max output: 256K
Modalities: Chat
Capabilities: Tool usePrompt Cache

About Step 3.5 Flash

Step 3.5 Flash is StepFun's open-weight sparse mixture-of-experts language model for reasoning, coding, and agent work. It holds 196.81B parameters but activates about 11B per token, which is what keeps generation fast. It is the text-only predecessor of Step 3.7 Flash, released under Apache 2.0, and it supports tool calling, prompt caching, and a thinking mode.

Where it works well

Agentic coding: StepFun's model card reports 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0.
Reasoning: the card lists 83.5% on GPQA Diamond for graduate-level science questions.
Fast generation: the card reports 100 to 300 tokens per second, from only 11B active parameters.
Tool calling and prompt caching are supported, which suits agent loops that resend a long prefix.

When to choose another model

StepFun's card says it tends to produce longer reasoning trajectories than competing models, which costs output tokens on simple questions.
The card notes weaker results in highly specialized domains and in extended multi-turn conversations.
It is a text model without image input; Step 3.7 Flash is the StepFun Flash model built with a vision encoder.

Getting started

Create API key
Create a key in Console, then use it with every model on the platform.

Send your first request

Copy the example for your language and run it against the endpoint.

Text to textMessages API

POST/v1/messages

API format:

curl https://api.tokenlab.sh/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: sk-xxx" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "step-3.5-flash",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Hello!"}
    ]
  }'

Pricing

The Verified price applies to Verified, which costs less on most models. The Official price is the model maker's published price and applies to the more reliable Official route. Auto bills the route that completes the request.

Default spec

per 1M tokens

Official price: Input $0.10 / Output $0.30 / Cache read $0.02
Verified price: —
Discount: —

Prompt cache pricing

Cache read

Official price: $0.02
Verified price: —
Discount: —

	Official priceper 1M tokens	Verified priceper 1M tokens	Discount
Default spec	Input $0.10 / Output $0.30 / Cache read $0.02	—	—
Prompt cache pricing
Cache read	$0.02	—	—

Usage & activity

Success rate is the share of requests that completed. Latency is how long a full response takes; P95 means 95% of requests finished within that time.

Usage & availability

Last 24 hours

Requests
Success rate
P95 latency
Total tokens

Model performance

Last updated Oct 21, 2026, 12:47 PM UTC

Metrics appear once privacy and data volume thresholds are met.

Data is based on aggregate user requests, excluding status checks.

Open in Console

Open Step 3.5 Flash in Console with a prompt ready to edit or send.

Help me try step-3.5-flash with a short message at /v1/messages. Show the reply, latency, and cost.

Use cases

Repository bug fixing
Point an agent at a failing test suite and let it read files, edit code, and rerun commands; the benchmark profile on the model card is built around this loop.
Competition-style math
Work through multi-step algebra, combinatorics, and proof sketches where a visible chain of reasoning matters more than a quick reply.
Terminal automation
Drive shell tasks such as setting up environments, parsing logs, and scripting migrations through tool calls that return command output.
Self-hosted reasoning parity
Prototype against the hosted model, then move to your own GPUs later, since the weights are published under Apache 2.0.

Prompt examples

The test test_parse_dates fails after my last commit. Read parser.py, find the cause, and give me the smallest patch plus the command to confirm it.

Prove that the sum of the first n odd numbers is n squared, then show a second proof using a geometric argument.

Write a bash script that scans a log directory for 5xx spikes over five-minute windows and prints the top three offending paths.

This model has conditional pricing. Monthly token totals alone cannot produce a reliable estimate; use the detailed pricing for the request specification, cache, and applicable time window.

FAQ

What is Step 3.5 Flash?

It is StepFun's open-source sparse mixture-of-experts model for reasoning and agentic tasks. It has 196.81B total parameters and activates roughly 11B per token. StepFun publishes it under Apache 2.0 on Hugging Face, so the same model can run hosted or on your own hardware.

Does Step 3.5 Flash support tool calling?

Yes. StepFun designed it for agent workloads: send function definitions in an OpenAI-style or Anthropic-style request and handle the returned tool calls as you would with any other model.

Does Step 3.5 Flash accept images?

No. It is a text model. StepFun's later Step 3.7 Flash adds a vision encoder, and Step 5 Preview takes images and video, so choose one of those for visual inputs.

How does it differ from Step 3.7 Flash?

Step 3.5 Flash is the earlier release and works on text. Step 3.7 Flash keeps the same size class and adds selectable reasoning levels and structured-output support. Pick 3.5 for text-only workloads that already work well.

Where does Step 3.5 Flash fall short?

StepFun lists three issues on the model card: long reasoning chains, instability when inputs differ from training data, and weaker results in specialized domains and long multi-turn chats. Test it on your own prompts before committing production traffic.

How much does Step 3.5 Flash cost?

On TokenLab, Step 3.5 Flash costs Input $0.10 / Output $0.30 per 1M tokens. The pricing table above shows the full breakdown. Rates depend on the billing unit, specification, and usage. Compare matching conditions in the model's detailed pricing; a single rate does not determine the total cost.

What are the context window and output limit of Step 3.5 Flash?

Step 3.5 Flash accepts up to 256,000 tokens of context and returns up to 256,000 tokens in one response.

Which endpoint should Step 3.5 Flash use?

Use https://api.tokenlab.sh/v1/messages for Step 3.5 Flash. The request example below shows the matching code shape.

Which operations does Step 3.5 Flash support?

Step 3.5 Flash supports Text to text. Select an operation above to see its endpoint and request example.