Agent backends
Power tool-using loops that need reasoning, code execution plans, and a long working context at a low token price.
Independent tool overview
Step-3.5-Flash at a glance
Step-3.5-Flash is StepFun's Apache-2.0 open-weight reasoning model for coding and agent workflows. Its sparse mixture-of-experts architecture has about 196.8B total parameters but activates roughly 11B per generated token, pairing a 256K context window with relatively low hosted-API pricing.
Visit the official Step-3.5-Flash site ↗
- Model type
- Sparse mixture-of-experts language model
- Total parameters
- 196.81B
- Active parameters
- About 11B per generated token
- Context window
- 256K tokens
- Primary uses
- Coding, reasoning, research, and agents
- License
- Apache 2.0
- Official API price
- $0.10 input / $0.30 output per 1M tokens
- Review date
- August 30, 2026
Overview
What Step-3.5-Flash is
Step-3.5-Flash is a text-generation foundation model optimized for tool use, coding, long-context work, and agent loops. It can be called through StepFun's OpenAI-compatible international API, used through supported third-party hosts, or self-hosted from published weights.
The model's efficiency comes from sparse activation, sliding-window attention, and multi-token prediction. Only a subset of experts runs for each token, but the deployment still has to store and serve a much larger 196B-class model. The 11B active-parameter figure should not be read as an 11B-model memory requirement.
StepFun has since published the newer Step-3.7-Flash, which adds multimodal capabilities and broader updates. Step-3.5-Flash remains available and can still make sense when an existing text, coding, or agent workload is already validated on it or when its current price and deployment support are preferable.
Use cases
Who Step-3.5-Flash is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Coding assistants
Support repository exploration, implementation, debugging, and terminal-oriented engineering workflows.
Long-document analysis
Process large text or code inputs within a published 256K-token context window, subject to quality testing at the intended length.
Open-weight deployment
Run or adapt published weights when a team has the substantial hardware and operations capacity required.
Cost-sensitive API workloads
Use the official hosted endpoint when its quality is sufficient and per-token economics matter more than access to the newest flagship model.
Capabilities
Core Step-3.5-Flash features
Sparse MoE architecture
Routes each token through eight of 288 routed experts plus a shared expert, activating about 11B parameters from the full model.
256K context
Uses a hybrid of sliding-window and full attention to support large codebases, documents, and accumulated agent context.
Multi-token prediction
Includes an MTP head intended to improve decoding throughput by predicting multiple tokens in a forward pass.
Agent and tool-use focus
Training and published evaluations emphasize coding, terminal work, browsing, research, and multi-step task execution.
OpenAI-compatible API
The international StepFun endpoint works with common Chat Completions clients using the model ID step-3.5-flash.
Open weights
Weights, technical material, deployment recipes, quantized variants, and integration examples are publicly available under Apache 2.0.
Tool integrations
Official guides cover agent and coding tools including OpenClaw, Claude Code, Codex CLI, Roo Code, Kilo Code, and others.
Process
How the Step-3.5-Flash workflow works
Step 1
Choose hosted or self-hosted
Start with the official API for fast evaluation; consider local or dedicated deployment only when privacy, customization, or scale justifies the infrastructure.
Step 2
Create a representative test set
Include actual codebases, tools, languages, context lengths, failure cases, and output constraints from the intended workflow.
Step 3
Connect through a supported protocol
Use StepFun's OpenAI-compatible endpoint or a documented integration, store the API key server-side, and pin the intended model ID.
Step 4
Implement tool boundaries
Validate arguments, sandbox code, restrict credentials and filesystem access, and require approval for consequential actions.
Step 5
Measure the full trajectory
Track not only token price but reasoning length, tool calls, latency, retries, completion rate, and human correction time.
Step 6
Compare with newer models
Re-run the same evaluations against Step-3.7-Flash and other current candidates before standardizing on an older model generation.
Cost
Step-3.5-Flash pricing and free plan
The model weights are available under Apache 2.0, but self-hosting still incurs hardware and operations costs. StepFun's international pay-as-you-go API lists Step-3.5-Flash at $0.10 per 1M uncached input tokens, $0.02 per 1M cached input tokens, and $0.30 per 1M output tokens. Step Plan subscriptions offer quota-based access for high-frequency agent and coding use.
Open weights
No model license fee
Download and deploy under Apache 2.0 while paying the resulting infrastructure and operations costs.
- Full and quantized weights are available
- 196B-class storage and memory requirements remain material
- Serving performance depends heavily on hardware and runtime
- Security, scaling, and monitoring are self-managed
Pay-as-you-go API
$0.10 input / $0.30 output per 1M tokens
Official hosted inference through StepFun's international platform.
- Cached input: $0.02 per 1M tokens
- No infrastructure to operate
- Rate limits depend on account tier and funding
- Actual cost depends on reasoning and output length
Step Plan Flash Mini
$6.99
Entry subscription for supported agent and coding integrations.
- 100 prompts per 5-hour window
- 400 prompts per week
- A prompt is a quota unit, not necessarily one API request
- StepFun estimates roughly 15–20 requests per prompt
Step Plan Flash Plus
$9.99
Higher quota for daily productivity.
- 400 prompts per 5-hour window
- 1,600 prompts per week
- Supports Step-3.5-Flash and the 2603 variant
- Quota conversion reflects context and tool usage
Step Plan Flash Pro
$29
Higher-volume plan for heavy agent use.
- 1,500 prompts per 5-hour window
- 6,000 prompts per week
- Designed for more intensive coding and agent workflows
- Plan availability and terms should be checked before purchase
Step Plan Flash Max
$99
Largest published Step Plan quota for professional workloads.
- 5,000 prompts per 5-hour window
- 20,000 prompts per week
- StepFun estimates up to roughly 300,000 requests per week
- Not a guarantee that every workload will reach the estimate
Pricing checked . Check current pricing at the source ↗
Assessment
Step-3.5-Flash strengths and limitations
Where it stands out
- Low official API pricing for a reasoning and agent-oriented model
- Apache-2.0 weights permit self-hosting and modification
- Large 256K context window for code and document-heavy tasks
- Sparse architecture activates a small fraction of total parameters per token
- Official guides cover common coding assistants and agent frameworks
- StepFun explicitly documents several known behavioral weaknesses
What to consider
- A 196B-class model remains demanding to store and serve even though about 11B parameters are active per token
- StepFun says the model can use longer reasoning trajectories than some peers, reducing the savings implied by headline token rates
- The vendor notes instability in specialized domains and long-horizon multi-turn conversations, including repetition, mixed languages, and time or identity inconsistency
- Published throughput and benchmark figures depend on hardware, serving configuration, harnesses, and vendor methodology
- It is a text model; the newer Step-3.7-Flash is the more relevant StepFun candidate when multimodal input is required
- Open weights do not provide a finished safety layer, agent sandbox, monitoring stack, or production service-level agreement
Compare
Step-3.5-Flash alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
GLM-5-Turbo
A newer high-speed agentic model positioned specifically for OpenClaw and similar long-running tool workflows.
Explore GLM-5-Turbo →Coding
Qwen3-Coder-Next
A more compact open-weight option designed around agentic coding when local deployment size matters.
Explore Qwen3-Coder-Next →Business Operations
DeepSeek
A broader open-model ecosystem with hosted chat and API options for reasoning, coding, and general-purpose workloads.
Explore DeepSeek →Questions
Step-3.5-Flash FAQs
Is Step-3.5-Flash open source?
StepFun publishes the weights and repository under the Apache 2.0 license. Teams should still review the license and any third-party dependencies or datasets relevant to their deployment.
What does 196B total and 11B active parameters mean?
The model contains about 196.8B parameters, but its router selects a subset of experts so roughly 11B participate in generating each token. That reduces computation; it does not shrink the full weight set to an 11B model.
Can Step-3.5-Flash run locally?
Yes, official full and quantized releases support local deployment, and StepFun gives examples using high-end consumer or compact AI hardware. Memory, storage, speed, and context capacity still depend on the quantization and machine.
How much does the StepFun API cost?
The international platform currently lists $0.10 per 1M uncached input tokens, $0.02 per 1M cached input tokens, and $0.30 per 1M output tokens for Step-3.5-Flash. Reasoning length can materially affect the effective cost of a task.
Does Step-3.5-Flash support a 256K context window?
Yes, the official specification lists 256K. Long-context support is a capacity claim, not proof that accuracy stays constant across every position or task, so evaluate it at the lengths you actually need.
Is Step-3.5-Flash good for coding agents?
Coding and agent workflows are its main positioning, and StepFun publishes integrations and benchmark results for them. Production selection should still use repository-specific tests, tool-call completion rates, security checks, and total task cost.
Should I use Step-3.5-Flash or Step-3.7-Flash?
Test both. Step-3.7-Flash is newer and multimodal, while Step-3.5-Flash remains a low-cost, text-focused option with established deployment recipes. The better choice depends on accuracy, modality, latency, compatibility, and cost in the actual workflow.
Bottom line
Our Step-3.5-Flash verdict
Step-3.5-Flash remains an attractive open-weight engine for cost-sensitive coding and agent experiments, especially through the inexpensive official API. Its sparse design is efficient but does not make self-hosting trivial, and StepFun's own known-issues section is a useful warning against trusting long, specialized agent runs without supervision. Evaluate it beside Step-3.7-Flash and current peers before committing a new production stack.
Visit Step-3.5-Flash website ↗Similar Consumer Tools
View all Consumer AI tools →.png)
Qwen3-Max-Thinking
Qwen3-Max-Thinking - Alibaba's new flagship reasoning model competitive with models like Claude 4.5 Opus, GPT 5.2-Thinking, and Gemini 3 Pro across benchmarks
Claude Opus 4.6
Claude Opus 4.6 - Anthropic's upgrade to its most powerful model line, featuring multi-agent collaboration, 1M context window, and new Office integrations
ERNIE 5.0
ERNIE 5.0 - Baidu's omni-modal, top-ranked Chinese model
Tiny Aya
Tiny Aya - Cohere Labs' new open-source multilingual small model covering 70+ languages in just 3.35B parameters
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.