Skip to main content

Step 5 Preview: 600B MoE, 27B Active — The Open-Source Powerhouse That Cuts Costs

Step 5 Preview: The Sparse MoE That Does More with Less

Step has just fully unveiled its new flagship base model, Step5Preview. It's built for real-world agent tasks, and it tackles a classic AI dilemma: the tug-of-war between capability, efficiency, and cost. Usually, improving one means sacrificing another. Step's answer? A sparse MoE architecture that expands the entire Pareto frontier.

Here's the kicker: Step5Preview has 600B total parameters, but only activates 27B at a time. It natively handles a 1 million token context window and processes both text and images. On the Artificial Analysis Intelligence Index, it scored 44 points, landing among the top three global open-source models. And the cost? Just one-eighth of Claude Opus5 per task. That means you get either stronger intelligence for the same budget or the same power for less.

Image

Hard Skills That Turn Heads

On tough benchmarks, Step5Preview flexes. It ranks second only to GPT-6Astra or Claude Opus5 on Agents' Last Exam (CLI subset ALE-CLI), financial research evaluation FrontierFinance, and cross-domain deep research DRACO. Other open-source models trail behind. GDPval-AA v2 also adopted the latest results as of September 19th, with DeepSWE v1.1 evaluated under the SWE-agent harness at temperature=1.0 and top_p=0.95.

In coding, Step built its own StepCodeBench, covering 553 independent repositories, 9 task types, 20 application areas, and 33 programming languages. Step5Preview shines in overall success rate and cross-scenario stability. It handles bug fixing, feature development, refactoring, and environment configuration. In one demo, it read a pile of ESP32 documentation, turned an ESP32-S3 board into a Vibe Coding keyboard with Bluetooth buttons and voice input. Along the way, it opened COM serial ports, adjusted cameras, captured device images, simulated mice, and kept debugging based on real errors — all over three hours.

Image

Long-Haul Tasks: Where It Truly Shines

Give it 24 hours and one H100 to optimize an MLA GPU kernel from scratch (head dimension 512, batch 1, 64 attention heads, 8192 tokens). Step5Preview independently modified code, ran the kernel, and measured throughput. If a solution worked but slowed things down, it dropped it and continued from the best version. After about 22 hours, it pushed peak performance to 508 TFLOPS, beating Claude Opus5's 493 TFLOPS.

Another 24-hour experiment asked it to improve a Qwen3-30B-A3B base model on AIME24 via automated post-training. The model decided how to use the annotated API and adjust data itself, raising accuracy from 53.3% to 60% — matching Opus5 while using fewer annotated tokens.

Beyond Code: Front-End, 3D, and History

Step5Preview doesn't just write web pages. It can call Blender to build and tweak 3D assets, then integrate them into Three.js for interactive apps and games. It even extracted info from a 1922 book, The Guangji Railway Travel Guide, to create a journey query, cost calculation, and route playback system — letting a small train run along historical routes. Talk about turning records into interactive products.

Finance: A Test of Comprehensive Ability

Step set up three internal evaluations around corporate research: FinStepBench-LiveSearch (retrieving and verifying financial info as needs change), CorporateValuation (turning data and assumptions into reproducible valuations), and DeepResearch (from evidence collection to a full report). The external benchmark FrontierFinance covers six investment scenarios with 220 expert questions and 11,543 scoring criteria. Step5Preview delivered strong results across all four.

The Bigger Picture

From Step3.5Flash to Step3.7Flash to Step5Preview, Step has consistently bet on one idea: the next stage of scaling isn't just about adding more compute — it's about using compute more efficiently. After today's release, the full model weights will be available on October 15th. This open-source flagship, which has recombined capability, cost, and efficiency, is ready for developers to make it their daily driver for agent tasks.

Key Points

  • Step5Preview uses a sparse MoE architecture: 600B total parameters, only 27B active.
  • Scores 44 on Artificial Analysis Intelligence Index, top three among open-source models.
  • Cost per task is one-eighth of Claude Opus5.
  • Supports 1M token context and multimodal (text + visual) input.
  • Excels in coding, long-horizon tasks, front-end/3D, and financial research.
  • Full model weights release on October 15th.