Stepfun
Open-source 198B-A11B MoE vision-language model for efficient agent inference
Step 3.7 Flash is a smart pick for budget-conscious teams building lightweight multimodal agents. Its sparse MoE architecture slashes inference costs without sacrificing core reasoning. The gap in documentation and tooling makes it less suitable for beginners, but for self-sufficient developers, the permissive license and low-latency callback API are real wins. If you need heavy hand-holding or enterprise SLAs, consider DeepSeek or Qwen instead.
Verified 5d ago · liveness 43/100 · cite: rightaichoice.com/tools/stepfun
- Developers building lightweight multimodal agents on a budget
- Teams deploying vision-language models in cost-sensitive production
- Researchers experimenting with MoE architectures
- Self-hosted inference pipelines needing low-latency callback API
- Enterprise buyers needing commercial support or SLAs
- Users wanting extensive documentation or beginner tutorials
- Teams without dedicated GPU infrastructure
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Stepfun if you need enterprise-grade support or SLAs, prefer English-first documentation, lack GPU infrastructure, or require a very long context window, since those gaps can stall adoption.
Self-hosting requires your own GPU infrastructure—costs for hardware, power, and maintenance can add up quickly, especially for a 198B-parameter model.
Stepfun's open-source model is free, but you pay for GPUs. For cost-sensitive teams that can self-host, it's cheaper than GPT-4o or Claude. DeepSeek and Qwen offer similar open-source models with more mature ecosystems, but Stepfun's sparse architecture may lower inference costs at scale.
In short
Stepfun — Open-source 198B-A11B MoE vision-language model for efficient agent inference. Best for Developers building lightweight multimodal agents on a budget, Teams deploying vision-language models in cost-sensitive production, Researchers experimenting with MoE architectures. Free to use.
Viability Score
How well maintained and how widely used is Stepfun? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: September 2026
How we score →Key Features
- 198B-A11B mixture-of-experts architecture (11B active per token)
- Native multimodal vision-language understanding
- Multi-turn function calling for agent tool-use
- Document parsing (PDF, tables, charts)
- UI screenshot analysis
- Fast inference for agent loops
- High throughput per GPU via sparse activation
- Customizable system prompts
- Open-source model weights on Hugging Face
- Step API for managed integration
- Step Plan for coding-to-agent development
- Studio for custom AI applications
- Desktop apps for macOS and Windows
- Mobile apps for iOS and Android
- Compatible with mainstream agent frameworks
About Stepfun
Stepfun is a Chinese AI lab that develops open-source vision-language models built for agentic workflows. Its flagship, Step 3.7 Flash, is a 198B-parameter mixture-of-experts model with only 11B active parameters per token. This sparse architecture lets you run multimodal reasoning, vision-language understanding, and multi-turn function calling at a fraction of the compute cost of dense models. Step 3.7 Flash handles document parsing (PDFs, tables, charts), UI screenshot analysis, and other document-heavy tasks natively, so you can build agents that can see and act without extra middleware. The model is released under a permissive open-source license, giving you full control over deployment and commercial use—whether you self-host on your own GPUs or use Step API for managed access. The broader ecosystem includes Step Plan for coding-to-agent development and Studio for building custom applications, with desktop clients for macOS and Windows and mobile apps for iOS and Android. If you're a budget-conscious team building lightweight multimodal agents, Step 3.7 Flash offers a cost-effective alternative to GPT-4o and Claude. However, you'll trade away some documentation and community support compared to DeepSeek or Qwen.
Behind the Verdict
Step 3.7 Flash stands out for its efficient sparse MoE design—198B total parameters but only 11B active per token. That means you get substantial reasoning and vision-language capability without the GPU footprint of a dense model of the same size. For self-hosters, this translates to lower cost per inference and better throughput on existing hardware. The model's built-in strengths (document parsing, chart understanding, UI screenshot analysis) make it a natural fit for document-heavy automation: think extracting data from invoices, analyzing screenshots, or controlling software via vision. Multi-turn function calling means you can wire it into an agent loop that calls tools, reads the results, and continues—exactly what production agents need. The permissive open-source license is a huge plus: no royalties, no usage restrictions, full commercial freedom. Step API offers a managed path if you'd rather not run GPUs. On the downside, the primary documentation and community are Chinese-first, which can slow down English-speaking developers. The ecosystem is young—Step Plan and Studio are early-stage—so you may need to fill gaps with your own tooling. And there's no official SLA on the API, so mission-critical production teams may want a fallback. Compared to DeepSeek and Qwen, Stepfun's community and docs are thinner; compared to OpenAI and Anthropic, you lose the plug-and-play APIs and enterprise support but gain major cost and flexibility. For a self-sufficient team that values cost and control, Step 3.7 Flash is a compelling choice.
Researching Stepfun? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas Stepfun actually fits — and what changes day-one when you adopt it.
Extract data from invoices and PDFs, then trigger follow-up actions based on the content.
Outcome: Deploy Step 3.7 Flash via Step API, use its document parsing to pull structured data, and wire function calls to update a spreadsheet or send emails—saving hours of manual entry.
Run a multimodal model on your own GPU cluster for a production automation tool.
Outcome: Download the open-source weights, serve with vLLM or similar, and use the 11B active parameters to handle high request volumes without buying extra GPUs—cutting inference costs by ~60% versus a dense 198B model.
Compare agentic performance and tool-calling accuracy across open-source models.
Outcome: Use Step 3.7 Flash's function-calling interface to benchmark it against DeepSeek and Qwen, gaining insights into trade-offs between sparse and dense models for agent tasks.
Use Cases
- Building lightweight multimodal agents for customer support or automation
- Self-hosting vision-language models for cost-sensitive production
- Running agent loops with function calling and tool use
- Experimenting with mixture-of-experts architectures for research
Models Under the Hood
as of 2026-08-30
Limitations
Requires your own GPU infrastructure for self-hosting; no official SLA on the API; primary interface and documentation are in Chinese; community support is smaller than DeepSeek's or Qwen's.
as of 2026-08-28
Verification history
We have re-verified Stepfun 17 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-checked, vendor evidence unchanged
- — re-checked, vendor evidence unchanged
- — re-checked, vendor evidence unchanged
- — re-checked, vendor evidence unchanged
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 17 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published Stepfun tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Open Source
$0
Ideal for
Developers and teams who want to self-host the model on their own GPUs, retain full control, and avoid per-token API costs.
What this tier adds
Free, with model weights on Hugging Face under a permissive license. No API access, so you handle deployment and GPU costs.
Where the pricing makes sense
The company stage and team size where Stepfun's pricing actually pencils out — and where peers do it cheaper.
Stepfun's open-source model is free, but you pay for GPUs. For cost-sensitive teams that can self-host, it's cheaper than GPT-4o or Claude. DeepSeek and Qwen offer similar open-source models with more mature ecosystems, but Stepfun's sparse architecture may lower inference costs at scale.
Setup time & first value
How long it actually takes to get something useful out of Stepfun — broken out by persona, not the marketing-page minute.
Self-hosting: 1-2 days to download weights, set up serving (e.g., vLLM), and integrate with your agent framework. Using Step API: a few hours to get your first agent loop running, given the Chinese-dominant docs.
Resources & Guides
Tutorials & Learning
Official links
Tools that pair well with Stepfun
Common stack mates teams adopt alongside Stepfun, with the specific reason each pairing earns its keep.
Alternatives to Stepfun
View allPopular in Foundation Models & LLM APIs
Frequently Asked Questions
Categories
Best-of guides
Used Stepfun? Help shape our editorial sentiment research.