StepFun's Step 5 Preview Uses 600 Billion Parameters — But Only 27 Billion at a Time
On September 20, 2026, Chinese AI lab StepFun announced Step 5 Preview, a 600-billion-parameter language model that, according to the company, matches much larger rivals on several benchmarks while running for roughly a third of the cost — StepFun is pricing input tokens at $1 per million (five cents on a cache hit) and output at $2.70 per million. The trick behind both the size and the price is architectural: Step 5 Preview is a sparse mixture-of-experts model, and only about 27 billion of its 600 billion parameters — under 5% — actually do any work on a given piece of text. Weights are due to open on October 15.
The problem: bigger models cost more to run, every single time
A conventional "dense" language model uses all of its parameters on every token it processes, whether that token is "the" or the name of an obscure enzyme. Make the model bigger to make it smarter, and you pay for that size on every single word it reads or writes — training cost and inference cost both scale with the full parameter count.
Mixture-of-experts (MoE) architectures — used by Step 5 Preview and by other large models from labs including DeepSeek, Mistral, and Google — break that link. The idea is to keep a huge pool of parameters available, but only activate a small slice of them for any given token.
How the routing actually works
Inside an MoE model, most layers aren't one big block of neurons — they're split into many smaller specialist sub-networks, usually called "experts." A separate, much smaller piece of the model, the router (or gating network), looks at each token as it passes through and decides which handful of experts are best suited to handle it, then sends the token only to those.
Each expert does its work independently, and their outputs are blended back together, weighted by how confident the router was in picking them, before moving to the next layer. The rest of the experts — the ones not selected for that particular token — sit idle. A different token later in the same sentence might get routed to a completely different set of experts.
The result is a model with a very large total parameter count (all those experts, combined) but a much smaller active parameter count on any given step — StepFun's reported 27-billion-of-600-billion split is one public example of that ratio. Because inference cost tracks the active count rather than the total, the model can afford a much bigger overall pool of knowledge without a proportionally bigger compute bill on every token.
Why "narrow-deep" matters too
Reporting on Step 5 Preview also describes it as using a 92-layer "narrow-deep" transformer layout — many layers, each comparatively slim — rather than fewer, wider layers. Combined with sparse MoE routing, this is part of a broader trend in 2026-era model design: separate how much a model knows (total parameters, spread across many experts and layers) from how much compute it spends per token (active parameters), and tune each independently instead of treating them as one and the same number.
Why it matters
Mixture-of-experts is a big part of why AI labs have kept shipping progressively larger models without inference costs exploding at the same rate — it's the same broad idea behind other MoE releases this year, including Alibaba's Qwen3.8-Omni-Flash. For anyone choosing between models on cost and capability, "total parameters" alone is no longer a reliable stand-in for "how expensive is this to run" — the active-parameter count, and the pricing that follows from it, is the number that actually predicts your bill. Whether Step 5 Preview's benchmark claims hold up against independent testing once its weights are actually open on October 15 is still an open question; for now, the pricing and architecture are what StepFun itself has published.
This explainer was made with LineLapse — type a topic, get a hand-drawn video. Make your own →