Skip to main content

60B Parameters, K3-Level Power: Step 5 Preview Opens to All

Step 5 Preview, a new large model from domestic AI firm JumpStellar, has officially opened its doors to everyone. The model first turned heads yesterday with a bold claim: it delivers the performance of a 2.8-trillion-parameter K3-level system using just 60 billion parameters. Today, the API and Studio are live, and a full-weight version is slated for release on October 15.

Under the Hood: Sparse MoE and a Million-Token Context

At its core, Step 5 Preview uses a sparse Mixture-of-Experts (MoE) architecture. It boasts a total of 600 billion parameters, but only activates 27 billion per task—a design that keeps costs down while maintaining serious muscle. It also supports a 1 million token context window and handles multi-modal input, meaning it can process text, images, and more.

Image

Performance: Matching K3, but Cheaper

On the AA Intelligence Index, Step 5 Preview scored 44 points, putting it on par with Kimi K3 Max. Its output speed clocks in at around 100 tokens per second, and the cost per task is just $0.71—compared to Kimi K3 Max's $2. That's a significant savings for similar performance.

Long-Haul Tasks: 24 Hours of Nonstop Optimization

Where Step 5 Preview really shines is in long-duration tasks. In one experiment, it autonomously optimized a set of H100 GPU kernels for 24 consecutive hours—writing code, running tests, comparing results, and iterating. After about 22 hours, it hit 508 TFLOPS, edging out Claude Opus 5's 493 TFLOPS in the same test.

In another 24-hour run, the model designed and trained its own data, boosting Qwen3-30B-A3B's accuracy on AIME24 from 53.3% to 60%—matching Claude Opus 5 while using fewer annotated tokens.

Free Access: Up to 75 Days on the House

JumpStellar is rolling out the red carpet for new users. Sign up and you'll get a 99 yuan package immediately. Log in and you'll snag 15 days of access; make your first call and you'll get another 15 days. Invite a friend to register, and you'll earn 15 more days per invite, up to a maximum of 45 days. All told, that's up to 75 days of free usage.

The Fine Print: Not Quite K3, But Close

Early feedback suggests Step 5 Preview doesn't quite match K3 across the board—but it wasn't built solely to chase benchmarks. Its overall capability is said to be close to DeepSeek V4.1 Flash. However, actual usage speed isn't blazing fast, and if free users flood in, things could slow down further.

Key Points:

  • Step 5 Preview uses a sparse MoE architecture with 600B total parameters, activating 27B per task.
  • It matches Kimi K3 Max on the AA Intelligence Index but costs $0.71 per task versus $2.
  • Long-task capabilities include 24-hour autonomous GPU kernel optimization, reaching 508 TFLOPS.
  • New users can get up to 75 days of free access through sign-up, login, first call, and referrals.
  • Real-world performance is close to DeepSeek V4.1 Flash, but speed may vary with user load.