Yesterday, a domestic large model called Step 5 Preview suddenly emerged, achieving the performance of a 2.8 trillion parameter K3-level model with just 60 billion parameters, offering very high cost-effectiveness. Today, the model has been fully opened—API and Studio are already available, and a full-weight model will be released on October 15th.
Step 5 Preview is the flagship model launched by JumpStellar, using a sparse MoE architecture, with a total of 600B parameters, activating only 27B each time, supporting a 1 million Token context and multi-modal input.
Benchmarking and Long Tasks: Autonomous GPU Kernel Optimization for 24 Consecutive Hours
In terms of performance, Step 5 Preview scored 44 points in the Intelligence Index on the AA list, matching Kimi K3 Max; its output speed is about 100 Token/s, with a cost of $0.71 per task, while Kimi K3 Max costs $2.
It also has good long-task capabilities: it can autonomously optimize a set of H100 GPU kernels for up to 24 consecutive hours, modifying code, running tests, comparing results, and continuing iteration. After about 22 hours, it reached 508 TFLOPS, while Claude Opus 5 achieved 493 TFLOPS in the same experiment. In another 24-hour experiment, Step 5 Preview designed and trained its own data, improving the accuracy of Qwen3-30B-A3B on AIME24 from 53.3% to 60%, matching Claude Opus 5, while consuming fewer annotated Tokens.
Free Package on Registration, Up to 75 Days in Total
JumpStellar has also launched an activity: new users get a 99 yuan package upon registration, 15 days upon login, and another 15 days after their first call. For each friend invited to register, users get 15 more days, with a maximum of 45 days in total, resulting in a total benefit of up to 75 days.
From current feedback, it's not easy for this model to truly match K3, but it's not designed solely for benchmarking. Its overall capability is close to DeepSeek V4.1 Flash. However, the actual usage speed isn't very fast, and if the number of free users increases, the speed may further decline.