Jieyue Star Open-Sources Step Code: Token Efficiency Matches DeepSeek Harness

iconKuCoinFlash
Share
AI summary iconSummary
Jieyue Star has open-sourced its AI programming tool, Step Code v0.1.0, under the MIT license, with growing interest as developers explore its potential. The tool runs in the terminal and supports code reading, writing, testing, and development tasks. It integrates MCP, Agent Skills, and Claude Code plugins. On Jieyue’s Terminal-Bench 2.1, it achieved an 80.9% task completion rate, matching DeepSeek Harness while using fewer tokens. In a 150-question Multi-Frame test, Step Code completed 110 tasks with a 73.3% score, using an average of 5.09 million tokens per question—the lowest among six tested harnesses. The Fear & Greed Index for AI tool development appears to be shifting toward optimism.
ME AI News, Dongcha Beating AI Brief: Jiepoin Star open-sourced Step Code v0.1.0, an AI-powered programming tool under the MIT license. It runs directly in the terminal, enabling code reading and writing, test execution, and development task automation, with support for MCP, Agent Skills, and Claude Code plugins. Its built-in StepPage allows publishing local static websites online with a single command. Step Code is designed to minimize token usage. In Jiepoin’s own Terminal-Bench 2.1 benchmark, it completed 72 out of 89 tasks with a score of 80.9%, tying with DeepSeek Harness for the highest score while using fewer tokens. On Jiepoin’s proprietary Multi-Frame long-task test with 150 questions, Step Code solved 110, achieving a 73.3% pass rate and averaging just 5.09 million tokens per question—securing both the highest pass rate and lowest token consumption among the six compared harnesses. However, Jiepoin did not disclose which underlying model was used in this benchmark or how the token savings were achieved, only stating that Step Code was optimized alongside the Step series models for reduced token usage. The source code reveals numerous “throttling” optimizations: file reads and command outputs are length-limited, searches minimize return ranges, long conversations default to context compression, and the system prompt encourages direct tool usage over invoking additional Agents. An even more aggressive contextProjection feature can trim old tool outputs, duplicate results, and historical reasoning—but it is disabled by default. Overall, its token efficiency stems from cumulative optimizations across both the model and the harness. (Source: BlockBeats)
Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information. Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.