Step 3.7 Flash: a 198B vision model that runs on desktop-class devices
Original title:A 198B vision model, running on a box that sits on a desk. This is what we built Step 3.7 Flash for....
AI summary
StepFun released Step 3.7 Flash, a 198B-parameter vision model designed to run on desktop devices such as the DGX Spark. User testing shows 128GB of unified memory is the threshold, with the model taking up about 104GB. Deployment does not require the official dedicated llama.cpp branch; mainline works. There are trade-offs in context length: with vision enabled, 64K on q8 KV cache is the ceiling; to use up to 256K context, vision must be disabled and switched to q4 KV cache, where model and cache together use about 114GB. It is a reasoning model, and the thinking process may consume large amounts of max_tokens, so settings need attention.
Key Takeaways
- StepFun released Step 3.
- 7 Flash,a 198B-parameter vision model designed to run on desktop devices such as the DGX Spark.
- User testing shows 128GB of unified memory is the threshold,with the model taking up about 104GB.
- Deployment does not require the official dedicated llama.
- cpp branch
Quick Facts
- Published
- Updated
- Source
- X:阶跃星辰 StepFun (@StepFun_ai)
- Category
- Tutorials & Tips
- AI score
- 80
把 198B 的视觉模型塞进一台桌面盒子,还跑通了,这本身就是个小里程碑。更关键的是,这篇实战直接帮你绕开了三个大坑,省下的三小时够你喝杯咖啡慢慢试了。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Step 3.7 Flash: a 198B vision model that runs on desktop-class devices. May 31, 2026. https://new.nat6.net/en/news/cmptg4k4t09zzsluzw0x9fbxd/