What is Step3-VL-10B?
Step3-VL-10B pairs a 1.8B-parameter vision encoder with a Qwen3-8B language model. It reads images and text, answers in text, and reasons step by step before it answers, with a 64K-token context in its standard mode. A second mode, PaCoRe, runs 16 reasoning passes in parallel and merges them, using up to 128K tokens, for harder problems.
StepFun publishes the weights but does not sell the model on its own API, so you run it yourself with vLLM, SGLang or Transformers. StepFun has released no newer small vision model since; its later Step 3.7 Flash also reads images but is a 198B model.
What are Step3-VL-10B's SeRe and PaCoRe modes?
Two ways to run the same weights: SeRe is the normal single reasoning pass, and PaCoRe spends far more compute for higher scores.
- SeRe
- The standard mode: one chain of reasoning with up to 64K tokens, and the mode behind most of StepFun's published scores.
- PaCoRe
- Runs 16 reasoning passes in parallel and combines them, with up to 128K tokens of context.
- The gain
- PaCoRe lifts MathVision from 70.8 to 76.0, MMMU from 78.1 to 80.1 and OCRBench from 86.8 to 89.0, by StepFun's figures.
- What you get
- The vLLM and SGLang set-ups in StepFun's quick start run the standard mode; PaCoRe multiplies the tokens and time each answer takes.
Who is Step3-VL-10B best for?
Step3-VL-10B is best for developers who self-host and need strong image reasoning, maths, OCR or screen reading from a model that fits on one 24 GB GPU.
Maths and diagram problems
StepFun reports 70.8 on MathVision in standard mode, against 52 to 60 for the other 7B to 10B models it tested.
Screen-reading agents
It scores 92.6 on ScreenSpot-V2 and 51.6 on ScreenSpot-Pro, StepFun's figures for finding buttons and fields in screenshots.
Documents and OCR on one GPU
It posts 86.8 on OCRBench and 89.4 on AI2D charts from a model that fits on a single 24 GB card.
Teams that must self-host
Apache 2.0 weights and an OpenAI-compatible endpoint through vLLM or SGLang keep your images on your own servers.
What's included with Step3-VL-10B?
Two reasoning modes, an OpenAI-compatible server through vLLM or SGLang, tool calling, FP8 and base weights, and high-resolution image crops.
- Two reasoning modes
- SeRe is a single reasoning pass with up to 64K tokens; PaCoRe runs 16 passes in parallel with up to 128K tokens.
- OpenAI-compatible server
- StepFun's vLLM and SGLang commands start a local endpoint that any OpenAI SDK can call.
- Tool calling
- The same serving commands switch on tool calls through the Hermes parser.
- FP8 and base weights
- Besides the chat model, StepFun publishes an FP8 version and a base model for fine-tuning.
- High-resolution images
- Each image is read as a 728 by 728 overview plus 504 by 504 detail crops.
How much does Step3-VL-10B cost?
Step3-VL-10B is free: the weights are a free download under Apache 2.0, and StepFun does not sell it on a paid API, so you pay only for the GPU you run it on.
- Chat, base and FP8 versions on Hugging Face and ModelScope
- Apache 2.0 licence
- Needs a GPU with at least 24 GB of memory
StepFun does not offer Step3-VL-10B on its paid API: its international and China price lists cover other models only. StepFun links a free browser demo on Hugging Face Spaces, and running the model yourself costs only your own hardware or rented GPU time. Full pricing on huggingface.co. We checked these prices on .
How do readers rate Step3-VL-10B?
No one has rated Step3-VL-10B yet. Be the first.
Sign in to rate. One rating per account.
Step3-VL-10B pros and cons
It scores well above other small vision models in StepFun's tests and fits on one GPU, but it always thinks before answering, reads images only, and has no hosted API.
Pros
- By StepFun's tests it leads other 7B to 10B vision models on most benchmarks, such as 78.1 on MMMU against 71.2 for GLM-4.6V-Flash, and early users on Reddit and Hugging Face call it one of the strongest small vision models.
- Fits on one 24 GB GPU such as an RTX 4090, and StepFun also publishes FP8 weights.
- Apache 2.0 licence with chat, base and FP8 weights, so you can fine-tune it and deploy it on your own terms.
- Local tools caught up after launch: llama.cpp and chatllm.cpp now run it, alongside vLLM and SGLang.
Cons
- It always reasons before answering: StepFun says it was not trained to skip thinking, and users who asked to turn it off or cap it got no official switch.
- Setup was rough at launch: several Hugging Face users hit vLLM errors that StepFun answered by pointing to nightly builds, and llama.cpp support came only later.
- Reads images and text only, with 64K tokens in standard mode, while GLM-4.6V and Qwen3.5 Small also take video and longer inputs.
- No hosted API from StepFun, and its highest scores come from the PaCoRe mode, which StepFun's quick start does not set up.
Is Step3-VL-10B open source?
Yes. Step3-VL-10B's code is public under the Apache 2.0 licence, so you can run it on your own server.
On GitHub
413stars
- Last commit
- 21 Jan
- Repository age
- 10 months
- Licence
- Apache 2.0
- Self-hosted
- Yes
- Repository
- stepfun-ai/Step3-VL-10B
What are the best Step3-VL-10B alternatives?
Step3-VL-10B is an open source alternative to Hunyuan Vision 1.5 Thinking. GLM-4.6V and Qwen3.6-27B do the same job and are open source too.
Hunyuan Vision 1.5 ThinkingPaidTencent's vision model for OCR, charts and visual reasoning on the TokenHub API.
GLM-4.6VOpen SourceChoose it when you want a hosted API or video input: GLM-4.6V-Flash is free on Z.ai's API, and the family reads video, images and files with a 128K context.
Qwen3.6-27BOpen SourceAlibaba's 27B open coding and vision model.
Qwen3.5 SmallOpen SourceChoose it when you need video input, a mode that answers without thinking, or a longer context: Qwen3.5-9B takes 262,144 tokens natively under the same Apache 2.0 licence.
Gemma 4Open SourceChoose it when memory is tight or you need audio: its 12B model reads images, video and audio, needs about 6.7 GB at 4-bit, and lets you configure thinking.
Questions about Step3-VL-10B
What is Step3-VL-10B?
Step3-VL-10B is StepFun's free, open 10B vision-language model for developers who want strong image reasoning, maths and screen reading on a single 24 GB GPU, under Apache 2.0.
Who is Step3-VL-10B for?
Step3-VL-10B is best for developers who self-host and need strong image reasoning, maths, OCR or screen reading from a model that fits on one 24 GB GPU.
Is Step3-VL-10B free?
Yes. Step3-VL-10B has a free Open weights plan: chat, base and FP8 versions on Hugging Face and ModelScope.
Is Step3-VL-10B open source?
Yes. Step3-VL-10B's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Step3-VL-10B alternatives?
The closest alternatives on Recaply are Hunyuan Vision 1.5 Thinking, GLM-4.6V, Qwen3.6-27B, Qwen3.5 Small and Gemma 4.
Where can I try Step3-VL-10B?
On its official website, stepfun-ai.github.io.
What GPU do I need to run Step3-VL-10B?
At least 24 GB of GPU memory, such as an RTX 4090 or A100: the weights take about 20 GB and running them adds about 4 GB. StepFun also publishes FP8 weights, and its Transformers code runs in bf16 only.
Does Step3-VL-10B have an API?
Not from StepFun. Its paid API price lists cover other models, so you serve Step3-VL-10B yourself: vLLM or SGLang turn it into an OpenAI-compatible endpoint on your own machine or a rented GPU.
Can I turn off thinking in Step3-VL-10B?
Not officially. StepFun says the model was not trained to answer without reasoning, so a chat template that skips thinking may behave unpredictably. StepFun said a non-thinking version would follow, but none is listed on its Hugging Face page.
Step3-VL-10B vs GLM-4.6V-Flash: which is better?
On StepFun's own tests Step3-VL-10B leads GLM-4.6V-Flash on MMMU (78.1 vs 71.2), MathVision and ScreenSpot-Pro. GLM-4.6V-Flash reads video, holds 128K tokens and is free on Z.ai's API, so it suits video and hosted use better.
Is Step3-VL-10B still StepFun's newest vision model?
It is StepFun's newest small open vision model. StepFun's later Step 3.7 Flash also reads images, but it is a 198B mixture-of-experts model that needs a server or a machine with at least 128 GB of unified memory.
Our verdict
Step3-VL-10B is the small open vision model to try first if you can host it yourself.
StepFun's published scores put it well ahead of other 10B-class models such as GLM-4.6V-Flash on reasoning, maths and screen reading, it fits on one 24 GB GPU, and Apache 2.0 lets you ship it.
- Skip it for
- Video work, hosted API use, or apps that need quick answers without a reasoning pass.
- Check first
- Test it on your own images, since the scores are StepFun's own and its best numbers use the slower PaCoRe mode.
Step 3.7 Flash
Perceptron Mk1
GPT-6.1 Terra
LFM3
Mistral Small 4
Step-3.5-Flash