OpenBMB releases JustGoal GPU kernels outperforming expert libraries on NVIDIA B200
OpenBMB released the JustGoal repository with GPU kernels generated by a general-purpose agent given only an optimization goal. On NVIDIA B200, JustGoal outperforms open-source libraries on 8 of 10 operators with per-operator geometric mean speedups of 1.14–2.34×. The code is Apache 2.0 licensed and available on GitHub.
China context
- Original name
- 面壁智能
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can directly use these Apache 2.0-licensed kernels to speed up inference on NVIDIA B200 GPUs, with reported per-operator speedups of 1.14–2.34× over open-source libraries.
- For investors
- The release demonstrates ModelBest's capability in AI-driven systems optimization, potentially reducing infrastructure costs for AI companies and creating a competitive advantage in efficient inference.
Translated from Chinese. Quotes and facts link to the original sources.
OpenBMB published the JustGoal repository, containing operator kernels from the paper 'JustGoal: With Only a Goal, the AI Baseline Is Above the Expert Kernels'. The kernels were produced by a general-purpose agent that received only an optimization objective and constraints, without kernel-specific evaluation infrastructure, knowledge bases, or search workflows. On NVIDIA B200, JustGoal outperforms open-source libraries on 8 of 10 operators, with per-operator geometric mean speedups of 1.14–2.34×, and surpasses the fastest publicly released specialized-agent kernel on 5 of 9 operators. The repository includes single-shape and full-shape goal runs, KDA-0.5 reproductions, and ablation studies, with environment details specifying NVIDIA B200 (sm_100a), CUDA 13.0, and CuTe DSL 4.6.2. The code is licensed under Apache 2.0, with some files retaining BSD-3-Clause headers from NVIDIA CUTLASS examples.
JustGoal demonstrates that a general-purpose agent, given only an optimization goal and constraints, can autonomously build an evaluation loop and iterate on kernel implementations to surpass expert-written libraries on NVIDIA B200. The repository provides kernels for multiple operators, including MLA paged prefill and GQA paged decode, with speedups measured against reference kernels. The KDA-0.5 harness reaches 0.95× the performance of JustGoal in a same-model comparison with Fable 5.1 on single shapes, indicating competitive performance with specialized-agent approaches.
GPU kernel developers and AI infrastructure teams outside China can now access Apache 2.0-licensed kernels that outperform open-source libraries on NVIDIA B200, potentially reducing the cost of optimizing inference for large models. The release of agent-generated kernels challenges the assumption that expert human tuning is required for peak performance, which may shift how companies allocate engineering resources for kernel optimization.
For companies running large-model inference on NVIDIA B200 hardware, adopting JustGoal kernels could reduce per-token serving costs by 1.14–2.34× on supported operators, directly improving margins for AI services. The open-source license allows commercial use without licensing fees, lowering the barrier to performance optimization.
The next verifiable signal is whether independent benchmarks on other B200 setups reproduce the reported speedups, as the repository welcomes retest results. Adoption by downstream projects such as FlashInfer or integration into production inference stacks would indicate practical impact beyond the paper's benchmarks.