600B Sparse MoE Activates 27B, Boosting Open-Source Flagship Cost-Effectiveness

Step 5 Preview, the latest flagship foundation model, is now officially available. Its purpose is clear: a primary model engineered for real-world Agentic tasks. It addresses the longstanding "Pareto problem" in AI, where capability, efficiency, and cost often conflict. Historically, progress in one area required sacrificing another. However, advancing the Pareto frontier isn't about removing trade-offs but expanding the entire boundary through innovative design.

This iteration utilizes a sparse Mixture of Experts (MoE) architecture. Step 5 Preview contains 600 billion parameters, with only 27 billion activated per inference. It natively supports a context window of up to one million tokens and processes both text and visual inputs. Scoring 44 points on the Artificial Analysis Intelligence Index, it ranks among the top three global open-source models. Notably, its single-task cost is just one-eighth that of Claude Opus5, offering either enhanced intelligence within the same budget or superior capabilities at a lower price. On ALE-CLI, a subset of Agents' Last Exam, as well as the financial investment research benchmark FrontierFinance and the cross-domain deep research evaluation DRACO, it trails only GPT-6Astra or Claude Opus5, leaving other open-source models behind. GDPval-AA v2 incorporates the latest data up to September 19th, while DeepSWE v1.1 was evaluated using the SWE-agent harness with temperature=1.0 and top_p=0.95.

In software development, Step created StepCodeBench, encompassing 553 independent code repositories, nine task types, 20 application domains, and 33 programming languages. Step 5 Preview demonstrates strong overall success rates and cross-scenario stability, effectively handling real-world tasks like bug fixing, feature development, refactoring, and environment configuration. In one demonstration, it processed extensive ESP32 documentation to transform an ESP32-S3 development board into a Vibe Coding keyboard with Bluetooth buttons and voice input. It autonomously opened COM serial ports, adjusted cameras, captured device images, simulated mouse movements, and iteratively debugged code based on real-time errors, completing the task over three hours.

Long-running tasks highlight its true value. In the first experiment, given 24 hours and an H100 GPU, it optimized an MLA GPU kernel from scratch (head dimension 512, batch size 1, 64 attention heads, 8192 tokens). Step 5 Preview independently modified the code, executed the kernel, and measured throughput. When encountering a solution that worked but reduced speed, it discarded it and continued from the best version. After approximately 22 hours, it achieved a peak performance of 508 TFLOPS, surpassing Claude Opus5's 493 TFLOPS. The second experiment also spanned 24 hours, allowing it to enhance a Qwen3-30B-A3B base model's performance on AIME24 through automated post-training. The model determined how to utilize the annotated API and adjust data, ultimately increasing accuracy from 53.3% to 60%, matching Opus5 while using fewer annotated tokens.

Regarding front-end and visual applications, it goes beyond writing web pages; it can invoke Blender to build and iteratively modify 3D assets, then integrate them into Three.js for interactive applications and games. It even extracted information from a 1922 "Guangjiu Railway Travel Guide," creating a journey query, cost calculation, and route playback system. This turned historical materials into an interactive product, featuring a small train running along the historical line.

Finance serves as a comprehensive capability test. Step developed three internal evaluations focused on company research: FinStepBench-LiveSearch tests the ability to retrieve and verify financial information as requirements evolve; CorporateValuation assesses converting data and assumptions into reproducible valuations; and DeepResearch evaluates the entire process from evidence collection to a complete research report. External benchmarks like FrontierFinance cover six investment scenarios with 220 expert questions and 11,543 scoring criteria. Step 5 Preview delivered strong results across all four benchmarks.

From Step3.5Flash and Step3.7Flash to Step5Preview, Step has consistently focused on one principle: the next phase of scaling is not merely about adding more Compute, but using Compute more efficiently. Following today's full release, complete model weights will be available on October 15th. This open-source flagship, which redefines the balance of "capability-cost-efficiency," is now ready for developers to adopt as their primary tool for Agentic tasks.