Stepfun
Open-source 198B-A11B MoE vision-language model for fast agent inference
A solid open-source multimodal model for agent-heavy workloads, undercutting comparable models on inference cost. Its ecosystem and documentation are still behind DeepSeek and Qwen, making it better suited for cost-savvy teams that can fill gaps themselves.
- Developers building lightweight multimodal agents on a budget
- Teams deploying vision-language models in cost-sensitive production
- Researchers experimenting with MoE models for agent and tool-use tasks
- Self-hosted inference pipelines needing fast callback latency
- Enterprise buyers requiring commercial support, API access, or SLAs
- Users needing extensive documentation and beginner-friendly tutorials
- Teams without dedicated GPU infrastructure for self-hosting
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 commercial API access, enterprise support, or extensive documentation.
GPU compute costs for self-hosting (not included)
Free open-source model; you pay only for your own GPU compute. Cheaper than proprietary APIs at scale, but requires upfront infrastructure investment.
In short
Stepfun — Open-source 198B-A11B MoE vision-language model for fast 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 models for agent and tool-use tasks. Free to use.
Viability Score
How likely is Stepfun to still be operational in 12 months? Based on 4 signals — momentum (how recently it shipped), wrapper dependency, revenue model, and web presence.
Last calculated: July 2026
How we score →Key Features
- 198B-A11B mixture-of-experts architecture
- Vision-language understanding and generation
- Fast inference for agent tool-use tasks
- Multi-turn function calling support
- Large context window for agent loops
- Open-source model weights on HuggingFace
- Strong multimodal reasoning and visual QA
- High throughput per GPU with sparse activation
- Customizable system prompts for agent behavior
- Low-latency callback API design
- Step-2 1.5B-step reasoning model on GitHub
About Stepfun
Stepfun is a Chinese AI lab behind Step 3.7 Flash, a 198B-A11B mixture-of-experts vision-language model released open-source in late May 2026. It activates only 11B parameters per token, enabling strong multimodal reasoning at low compute cost. Designed for agentic workflows, it excels at vision-language tasks, function calling, and multi-turn agent loops. The model is available under a permissive open-source license on Hugging Face. Key features include vision-language reasoning, multi-turn function calling, customizable system prompts, fast inference with low latency, and high throughput per GPU. Stepfun is pursuing a Hong Kong IPO, signaling growing commercial ambitions. Its ecosystem lags behind DeepSeek and Qwen in documentation and tooling, but Step 3.7 Flash is a cost-effective choice for developers prioritizing efficiency and permissive licensing over support.
Behind the Verdict
Step 3.7 Flash is for developers who need multimodal reasoning on a budget. The sparse MoE architecture delivers high throughput with only 11B active parameters, translating to lower GPU costs per query. It's particularly strong in agent loops and tool use, where it rivals larger models on benchmarks. But the trade-offs are real: community resources are thin, and documentation is sparse compared to the Qwen or DeepSeek ecosystems. There's no commercial API or SLA, so teams need in-house ops expertise. For researchers prototyping on a low budget, it's a great starting point. For enterprise deployments needing support, look elsewhere. The IPO filing suggests the lab is maturing, but don't expect enterprise features soon.
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.
Build a personal AI assistant that can answer questions about images and execute tool calls.
Outcome: Deploy Step 3.7 Flash on a single NVIDIA GPU using vLLM, achieving low-latency multimodal conversations with function calling.
Experiment with MoE architectures for agentic tasks on a limited budget.
Outcome: Fine-tune Step 3.7 Flash on custom agent data, leveraging sparse activation to reduce compute costs while maintaining performance.
Integrate a cost-effective vision-language model into a production agent pipeline.
Outcome: Self-host Step 3.7 Flash with Docker and CUDA, achieving high throughput per GPU for real-time agent inference.
Use Cases
- Building lightweight multimodal agents for customer support or automation
- Self-hosting vision-language models for cost-sensitive production pipelines
- Running agent loops with function calling and tool use
- Experimenting with mixture-of-experts architectures for research
Models Under the Hood
as of 2026-07-06
Limitations
Open-source model requires own GPU infrastructure; no official API or SLA; primarily Chinese interface and documentation; community support is smaller than some competitors.
as of 2026-06-26
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 with GPU infrastructure who want to self-host a free multimodal model for agent tasks.
What this tier adds
Free model weights on Hugging Face; no API or support included.
Where the pricing makes sense
The company stage and team size where Stepfun's pricing actually pencils out — and where peers do it cheaper.
Free open-source model; you pay only for your own GPU compute. Cheaper than proprietary APIs at scale, but requires upfront infrastructure investment.
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.
For a developer with GPU infrastructure: ~1-2 hours to download model weights from Hugging Face and run inference with vLLM. Fine-tuning or custom agent integration may take days.
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 proprietary API (OpenAI, Anthropic): Download open-source weights and self-host, removing per-token costs.
- →From other open-source models (e.g., Qwen-VL): Swap model files and adjust prompts; minimal code changes if using Hugging Face Transformers.
- ↗To another open-source model (e.g., DeepSeek-VL): Download new weights and update inference code.
- ↗To a proprietary API: Sign up for API access and replace local inference calls with API requests.
Integrations
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 allFrequently Asked Questions
Categories
Used Stepfun? Help shape our editorial sentiment research.