On February 2, Chinese language AI firm StepFun formally launched Step 3.5 Flash, its newest and strongest open-source basis mannequin, positioning it as “born for brokers.” The corporate says the mannequin delivers robust reasoning capabilities, excessive stability, and efficiency optimized particularly for Agent-based workflows.
Key highlights embody:
- Sooner: Inference speeds of as much as 350 tokens per second for single-request coding duties
- Stronger: Efficiency in Agent situations and mathematical reasoning corresponding to main closed-source fashions
- Extra secure: Able to dealing with advanced, long-horizon, multi-step duties
In response to StepFun, the following technology of basis fashions should not solely be “smarter,” but in addition reliable, responsive, and cost-efficient. To attain this steadiness, Step 3.5 Flash adopts a number of architectural improvements:
- Sparse Combination-of-Specialists (MoE): Every token prompts solely round 11 billion parameters, out of a complete 196 billion, considerably decreasing compute prices
- MTP-3 (Multi-Token Prediction): The mannequin predicts three tokens per step, successfully doubling inference effectivity
- Hybrid Consideration Structure (SWA + Full Consideration): A 3:1 sliding-window-to-global consideration combine permits the mannequin to deal with key segments in lengthy texts, enabling environment friendly processing of as much as 256K context size with decrease computational overhead
Step 3.5 Flash is now absolutely obtainable, and StepFun additionally revealed that coaching for the Step 4 mannequin has already begun. The corporate invited builders and researchers to take part within the mannequin’s open improvement and ecosystem co-creation.
With its emphasis on Agent intelligence, long-context reasoning, and inference effectivity, Step 3.5 Flash alerts StepFun’s ambition to ascertain a aggressive open-source basis for next-generation AI agent methods.
Supply: IT House
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 tendencies at the moment: learn extra, subscribe to our e-newsletter, and turn out to be a part of the NextTech neighborhood at NextTech-news.com
