GitLab expands AI model options for optimized cost-performance
New hosted open weight models offer 4x more calls per credit than frontier models
1Diverse AI model needs
Different software development tasks require varying AI model capabilities. Complex refactors need large context windows while routine tasks benefit from cost-efficient models. Teams often compromise by standardizing on one model due to governance challenges in regulated environments. This leads to inefficient resource allocation where high-cost models handle tasks better suited for cheaper alternatives.
2New open weight models
GitLab introduces three hosted open weight models: Kimi K3, GLM 5.3, and MiniMax M3. These complement existing frontier models, offering teams more control over quality, latency, and cost optimization. Each model is specialized - Kimi K3 for hard tasks, GLM 5.3 for long-horizon coding, and MiniMax M3 for high-volume work. The platform now provides up to 4x more calls per GitLab Credit compared to some frontier models.
3Performance benchmarks
Internal testing shows Kimi K3 delivers 1.82 calls per credit, nearly matching the default model's 2 calls. GLM 5.3 achieves 5 calls per credit while maintaining strong performance. MiniMax M3 leads in cost efficiency with 8 calls per credit. These options allow precise matching of model capability to task complexity, significantly reducing inference costs for agentic workloads.
4Model specialization
Kimi K3 excels with large codebases and extended agent sessions. GLM 5.3 maintains context across long coding tasks. MiniMax M3 optimizes for high-volume, multi-step work. This specialization lets teams avoid overpaying for capabilities they don't need while ensuring adequate performance where it matters most.
5Centralized governance
GitLab provides built-in control over model selection. Admins can set default models per feature and restrict available options across all child groups. The platform supports GitLab-managed, self-hosted, or hybrid deployments to meet security and compliance requirements. All models undergo rigorous vendor assessments including SOC 2 Type 2 and ISO 27001 verification.
6Security assurance
GitLab-managed models run entirely within GitLab's environment with zero data retention. Fireworks AI, hosting the open weight models, meets strict security standards including penetration testing and vulnerability remediation. This governance framework enables teams to adopt new models with confidence comparable to existing options.
7Implementation flexibility
Teams can select open weight models as optional alternatives or set them as defaults for specific features. The hybrid deployment option allows mixing GitLab-managed and self-hosted models based on workload requirements. This flexibility helps organizations balance performance needs with infrastructure constraints.
8Strategic impact
The expanded model roster transforms AI from a one-size-fits-all solution to a precision tool. Teams can now optimize both performance and cost at the task level, particularly valuable for long-running agentic workflows. This advancement helps bridge the gap between AI-assisted development speed and traditional review cycles.
Original article: GitLab — GitLab expands AI model options for optimized cost-performance