Stepfun
StepFun's OpenAI-compatible API bundles frontier reasoning, vision-language parsing and a full speech stack — realtime voice, ASR, TTS and
StepFun earns its place when you need voice, vision and document parsing without stitching together four vendors — StepAudio 3 plus Step 3.7 Flash covers that in one key, and the OpenAI-compatible endpoint keeps migration cheap. Go in expecting to be self-sufficient: the English docs and community pool are thinner than the US labs. Great fit for teams that already know what they're building, weak fit for anyone learning on the way.
Verified 1d ago · liveness 68/100 · cite: rightaichoice.com/tools/stepfun
- Developers building voice agents that also need vision and document parsing
- Teams migrating from OpenAI that want a base-URL and model-name swap
- Product builders assembling realtime voice, ASR, TTS and music from a single vendor
- Self-hosting teams that need downloadable weights for data control
- Beginners who depend on extensive tutorials and a large community answer pool
- Teams with no GPU infrastructure that still want to self-host the open weights
- Buyers who need a managed self-host product rather than raw weights
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 your team has no GPU capacity and no appetite to read model weights and raw API docs — the surrounding tutorials and community answers are thinner than DeepSeek's or Qwen's, so you will be largely on your own.
StepFun publishes three routes rather than one rate card: open weights at no cost if you supply the GPUs, a usage-based API for pay-as-you-go integration, and a Flash Pro monthly subscription that bundles coding, development and agentic workflows. The self-host path shifts spend from per-token billing to GPU capacity and engineering time — cheaper at volume, more expensive if you never reach volume.
In short
Stepfun — StepFun's OpenAI-compatible API bundles frontier reasoning, vision-language parsing and a full speech stack — realtime voice, ASR, TTS and. Best for Developers building voice agents that also need vision and document parsing, Teams migrating from OpenAI that want a base-URL and model-name swap, Product builders assembling realtime voice, ASR, TTS and music from a single vendor. Free to use.
No head-to-head comparison for Stepfun yet — the alternatives below are the closest matches we list.
What people actually say about Stepfun — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
20 mentions across 5 sources (Hacker News, YouTube, Product Hunt, App Store, Lemmy), 50 more we could not attribute · researched Sep 25, 2026.
Weighted by the 70 posts each of 5 sources contributed.
- +198B-A11B MoE runs at 11B active params per token, making frontier-class inference dramatically cheaper.
- +Users report ~$0.10/M input and $0.30/M output pricing that undercuts dense frontier models.
- +Local runners can own the weights and get uncensored, unrestricted answers StepFun-controlled models elsewhere refuse.
- +Step 5 Preview lands on the Artificial Analysis Pareto frontier, confirming genuine frontier-tier capability.
- +1M context window and video input in Step 5 Preview open long-document and video agent use cases.
- −Documentation and community support are visibly thinner than DeepSeek and Qwen.
- −iPad app is missing and fails to log in on iPadOS 18, blocking tablet users.
- −Local self-hosting struggles on non-NVIDIA hardware, with some users stuck below double-digit tokens/sec.
- −Step 5 Preview weights aren't actually released yet — the flagship is still a promise until Oct 15.
- −Model has been observed claiming to be Claude, raising identity-reliability concerns for branded agents.
- • GPU capex or cloud GPU rental if you self-host — hardware sensitivity means non-NVIDIA setups may need more or newer cards than expected.
- • Third-party provider markups (e.g. SiliconFlow) versus direct Step API pricing.
- • Engineering time to wire the model into mainstream agent frameworks since there is no native integration marketplace.
- • Step Plan's ~$60 tier is comparatively steep versus the model's cheap token economics.
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: October 2026
How we score →Key Features
- Step 5 Preview: frontier reasoning model for complex and agentic tasks
- Step 5 Preview weights downloadable from Hugging Face and GitHub
- Step 3.7 Flash: sparse mixture-of-experts vision-language model
- Document parsing for PDFs, tables and charts
- UI screenshot analysis for visual agents
- Step 3.5 Flash: lighter model tier for simpler workloads
- StepAudio 3 Realtime: real-time voice conversation
- StepAudio 3 ASR: speech recognition
- StepAudio 3 TTS: text-to-speech synthesis
- StepAudio 3 Gen: voice generation
- StepAudio 3 Music: music generation
- OpenAI-compatible POST /v1/chat/completions endpoint with Bearer auth
- AI Studio for in-browser model testing before writing code
- Self-hosting via open model weights with a permissive licence
- Vision and image understanding
About Stepfun
StepFun is a Shanghai-based AI lab selling model access through an OpenAI-compatible API, with model weights published openly under a permissive licence for teams that would rather self-host. Its work splits across two tracks: frontier reasoning and vision-language on one side, a full voice family on the other. On the frontier side, Step 5 Preview is the model StepFun pitches at complex work and agentic tasks. Its weights are now available on Hugging Face and GitHub, so the download option arrived rather than being teased. Step 3.7 Flash is a sparse mixture-of-experts vision-language model aimed at agent inference and document-heavy jobs — PDF, table and chart parsing, plus UI screenshot analysis for visual agents. Step 3.5 Flash covers lighter workloads. The voice track, StepAudio 3, is where the catalogue gets unusual. Realtime handles live voice interaction, ASR covers speech recognition, TTS covers synthesis, Gen does voice generation and Music does music generation. Voice-agent teams normally buy that spread from three or four vendors; here it is one family behind one Bearer token. The endpoint is POST https://api.stepfun.ai/v1/chat/completions, so an existing OpenAI app mostly needs a base-URL and model-name swap. Buyers can explore through AI Studio in the browser before writing code, use the API Platform for production, and subscribe to Flash Pro, a monthly plan bundling coding, development and agentic workflows. Against DeepSeek and Qwen, the case rests on modality breadth rather than a benchmark crown.
Behind the Verdict
We'd reach for StepFun when a project needs realtime voice plus document understanding under one roof. Assembling Realtime, ASR, TTS and music from separate vendors means four contracts, four auth schemes and four sets of latency characteristics to reconcile. One Bearer token across the family removes that plumbing, and the OpenAI-compatible endpoint means an existing chat app ports with a base-URL and model-name change rather than a rewrite. The second reason is self-hosting. Step 5 Preview weights are downloadable now from Hugging Face and GitHub under a permissive licence, which matters for regulated teams that cannot ship prompts to a third-party endpoint. In practice that shifts cost from per-token billing to GPU capacity and the ops staff to run it. Where it bites: documentation depth. The English docs and community answer pool are smaller than OpenAI's or Anthropic's, so debugging an odd ASR result or a bad Realtime handoff often means reading the spec yourself. Budget for that before committing a small team to it. Pick StepFun over DeepSeek or Qwen when the job is multimodal — voice agents that also need vision, or document pipelines that need both. If you only need text reasoning at the lowest price, a narrower specialist is easier to justify. If you need managed hosting rather than raw weights, the open-source route is not that product. One caveat on the roadmap: StepFun still labels Step 5 as Preview, so treat it as production-adjacent, not production-proven, for anything latency-critical.
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.
Swap the base URL to https://api.stepfun.ai/v1/chat/completions, authenticate with a Bearer token, and change the model name to step-3.7-flash while keeping the existing messages payload.
Outcome: A working model swap with no rewrite of the request layer, after which you can A/B the two providers on the same prompts.
Wire StepAudio 3 ASR for speech recognition, Step 3.7 Flash for reasoning over the transcript and attached documents, and TTS or Gen for the spoken reply.
Outcome: One API key covers listening, reasoning and speaking, instead of three vendor contracts and three billing systems.
Pull the open model weights from Hugging Face, deploy them on your own hardware, and keep the hosted API only for models or modalities you cannot run locally.
Outcome: Data stays in your environment and per-token spend converts into fixed GPU capacity you already own.
Use Cases
- Building voice-enabled agents that also need vision and document understanding from one API
- Parsing PDFs, tables, charts and UI screenshots in an agent pipeline
- Migrating an OpenAI-based app to StepFun models with a base-URL and model-name swap
- Self-hosting open-weight models for data control or cost reduction
- Assembling realtime voice, ASR, TTS and music without multiple vendors
- Running agent loops with function calling and tool use
- Experimenting with mixture-of-experts vision-language architectures
Models Under the Hood
as of 2026-08-30
Limitations
- Self-hosting the open weights requires your own GPU infrastructure and the operational work that comes with it.
- There is no published API SLA.
- The company is Shanghai-based and much of the surrounding documentation and community discussion is in Chinese, which slows debugging if your team works in English.
- StepFun's third-party community — forums, Stack Overflow answers, tutorials — is smaller than what surrounds DeepSeek or Qwen, so unusual failures are likely to land on your own engineers.
- The public product surface is narrow: AI Studio, the API Platform and the Flash Pro subscription, with no broad integrations marketplace to lean on.
as of 2026-09-29
Verification history
We have re-verified Stepfun 21 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-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
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-checked, vendor evidence unchanged
- — re-checked, vendor evidence unchanged
- — re-checked, vendor evidence unchanged
Showing the 6 most recent of 21 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
ML engineers and self-hosting teams with their own GPU capacity who need data control or lower per-token economics at volume.
What this tier adds
Starting point for the self-host route: open model weights you deploy and run yourself.
API Platform
Usage-based
Flash Pro
Monthly subscription
Ideal for
Individual developers and small teams working across coding, development and agentic workflows who want those bundled rather than metered.
What this tier adds
A monthly subscription that packages coding, development and agentic workflows together instead of billing purely by usage.
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 publishes three routes rather than one rate card: open weights at no cost if you supply the GPUs, a usage-based API for pay-as-you-go integration, and a Flash Pro monthly subscription that bundles coding, development and agentic workflows. The self-host path shifts spend from per-token billing to GPU capacity and engineering time — cheaper at volume, more expensive if you never reach volume.
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.
A base-URL and model-name swap against the OpenAI-compatible endpoint is roughly an hour for a developer with an existing OpenAI integration. AI Studio exploration is near-instant, no setup. Self-hosting the open weights is days, not hours — it depends entirely on your GPU capacity and deployment tooling. Flash Pro for coding and agentic workflows sits between the two.
Switching to or from Stepfun
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From OpenAI: point the client at https://api.stepfun.ai/v1/chat/completions, keep the same messages payload, and swap in a StepFun model name.
- →From another OpenAI-compatible provider: same base-URL and model-name change, since the request shape is already identical.
- →From a split voice stack: consolidate ASR, TTS, realtime voice and music onto StepAudio 3 under one API key.
- →From a closed hosted model: move to self-hosted StepFun weights when data residency or unit economics require it.
- ↗To a self-hosted deployment: take the open weights and your existing prompts, since the model family stays the same.
- ↗To another OpenAI-compatible provider: the reverse base-URL and model-name swap, with no change to your request layer.
- ↗To a managed voice vendor: replace StepAudio 3 with a dedicated speech provider if you need one that is not tied to a model lab.
Resources & Guides
Tutorials & Learning
YouTube returned 6 videos for “Stepfun”, and we withheld 6: 6 could not be judged, because “Stepfun” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Stepfun.
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 all
Popular in Foundation Models & LLM APIs
Frequently Asked Questions
Categories
Best-of guides
Used Stepfun? Help shape our editorial sentiment research.