Intel / post

Step 3.5 Flash 2603 cuts token spend with a low reasoning mode

Source: x.com
Date: Published Apr 2, 2026

StepFun@StepFun_ai

Step 3.5 Flash 2603 just shipped! Low mode: 56% fewer tokens High mode: full reasoning, still 14% leaner than before Agentic workflows make hundreds of calls—most don't need deep thinking. Not every call needs to be a philosophy session. Now you can dial it down. Live for all Step Plan users. https://t.co/lmZtnHwSTb

Apr 2, 2026 View on X

Context

StepFun's update to Step 3.5 Flash, labeled the 2603 checkpoint, adds a low reasoning mode that the company says uses 56% fewer tokens than before, alongside a high mode still tuned for full reasoning but 14% leaner than the prior version. Step 3.5 Flash is StepFun's model, launched with 196 billion total and 11 billion active parameters and aimed at and coding workloads.

StepFun's stated reason for the low mode is that agentic workflows make hundreds of API calls, and it argues most of those calls do not need deep reasoning, so developers can now dial reasoning effort down per call instead of always paying for full reasoning depth. The 56% and 14% figures are StepFun's own reported comparisons against the model's earlier checkpoint; no independent accompanies this specific update.

Terms in this piece · Glossary

open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source x.com