Chinese language large-model startup StepFun has totally open-sourced Step 3.5 Flash, releasing not solely the mannequin itself but in addition its Base (pretrained) weights, Midtrain weights, and the accompanying Steptron coaching framework—additional advancing its dedication to open AI infrastructure for agent growth.
Step 3.5 Flash adopts a sparse MoE (Combination-of-Consultants) structure with 196 billion complete parameters, whereas activating solely round 11 billion parameters throughout inference. Beneath single-request coding duties, inference pace can reportedly attain as much as 350 tokens per second (TPS). The mannequin is purpose-built for agent situations, demonstrating sturdy efficiency in complicated reasoning and long-horizon process chains. In accordance with the corporate, its reasoning depth rivals sure top-tier closed-source fashions.
The mannequin has gained fast traction throughout the open-source group. Downloads on Hugging Face have surpassed 300,000, and it has climbed to No.1 on the OpenRouter Trending榜. On the well-known open-source benchmark challenge OpenClaw (nicknamed “Little Lobster” by Chinese language builders), Step 3.5 Flash has risen to the highest two rankings.
With the discharge of full coaching weights and framework help, StepFun is positioning Step 3.5 Flash as a foundational open base mannequin for next-generation AI brokers.
Supply: AI Each day
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the newest breakthroughs, get unique updates, and join with a worldwide community of future-focused thinkers.
Unlock tomorrow’s traits as we speak: learn extra, subscribe to our publication, and turn into a part of the NextTech group at NextTech-news.com
