StepFun Opens Step 3.5 Flash to All

StepFun just opened Step 3.5 Flash to everyone—weights and all.

StepFun, a Chinese AI startup, released the entire Step 3.5 Flash package: the model, its Base weights, Midtrain weights, and the Steptron training framework. The project uses a sparse Mixture-of-Experts architecture with 196 billion total parameters, but only about 11 billion are active during inference. In single-request coding tasks, the model reportedly runs up to 350 tokens per second. The release positions StepFun as a rare example of a Chinese foundation model being fully open—weights, training tooling, and deployment guidance included.

The numbers aren’t just marketing. Sparse MoE (稀疏MoE) design lets StepFun keep the parameter count enormous while lighting up a tiny, relevant subset of experts for any given task. That, in theory, translates to strong reasoning and long-horizon task chaining without prohibitive compute costs. StepFun frames Step 3.5 Flash as an agent-ready base: not just a language model, but a platform for building AI agents that can plan, reason, and act across chained tasks. The company’s rhetoric is reinforced by traction signals on the open-source side: downloads on Hugging Face exceeded 300,000, it sits at No.1 on OpenRouter’s Trending榜, and in the OpenClaw benchmark project—known as “Little Lobster” to Chinese developers—StepFun’s model is among the top two.

What’s happening here, in plain terms, is a calculated bet on open AI infrastructure as a driver of domestic capability. By pairing the model with Base and Midtrain weights and the Steptron training framework, StepFun removes several traditional choke points: you don’t just get a model—you get the data, fine-tuning recipes, and tooling to tailor agents for specific industrial tasks. In a landscape where many leading AI systems are locked behind vendor APIs or opaque licensing, an open package of this scale lowers the barrier for Chinese robotics vendors, research groups, and enterprise SOCs to prototype and deploy AI-enabled agents on real factory floors.

From a manufacturing and supply-chain lens, the implications are tangible. StepFun’s release could accelerate the adoption of AI-powered agents for scheduling, predictive maintenance, and robotic orchestration across Chinese plants and supplier networks. Chinese servo and automation component makers, often serving global OEMs, can now integrate a fully open foundation model into their control software without paying ongoing API fees or negotiating bespoke licenses. That matters in a sector already wrestling with margin pressure and the need to validate AI safety and reliability on the shop floor.

The broader context matters too. Open-sourcing a large, capable model underpins what many in China describe as “自主可控” (self-reliant control) in AI infrastructure. It’s a dovetail to policy signals that encourage domestic innovation ecosystems, while also sharpening competitive tension with global providers. However, it brings caution: hosting, tuning, and validating such models for factory use requires robust data governance, safety testing, and alignment with industrial standards. The economics—large upfront development costs versus long-run open-source productivity—will hinge on how effectively StepFun’s base and mid-training weights, plus the Steptron framework, translate into reliable, auditable agent behavior on real equipment.

Two practitioner notes to watch:

Model engineering discipline matters as much as raw capability. The 196B parameter canvas sounds impressive, but real-world robot agents must handle sensor noise, PLC latency, and safety constraints. Expect growing interest in domain-adaptive fine-tuning and robust evaluation suites for manufacturing contexts.

Open-source cadence changes negotiation dynamics with OEMs. With weights and tooling public, the value shifts toward interoperability and support ecosystems. Vendors will weigh licensing, data-privacy considerations, and update paths more publicly, which could accelerate or complicate deployments across complex supply chains.

What this means for sourcing and competition is nuanced. For companies buying capacity in China, this move widens the set of locally adaptable AI options and may dampen the premium on external AI APIs for routine agent tasks. For incumbents outside China, StepFun’s openness presents a potential route to seed competing agents that align with Chinese standards, while raising the bar on what “open” means in a geopolitically sensitive AI era.

In short, StepFun’s 3.5 Flash release isn’t just a release—it’s a signal about how the next wave of industrial AI may be built, tested, and deployed in China: open, scalable, and deeply integrated with the factories that keep the world’s largest manufacturing ecosystem humming.

Sources

StepFun Fully Open-Sources Step 3.5 Flash