JetBrains Mellum

JetBrains Mellum: The PSI-Integrated Local Intelligence

As of January 13, 2026, Mellum has solidified its role as the industry standard for secure, on-device developer assistance. The 2026 architecture leverages a Sparse Mixture-of-Experts (MoE) design, which allows it to maintain the reasoning capabilities of a 12B model while only activating 2B parameters for specific tasks (e.g., Java vs. Kotlin refactoring), ensuring no thermal throttling on modern developer laptops 📑. The core technical advantage remains its native integration with the Program Structure Interface (PSI), enabling compiler-grade symbol resolution that outperforms standard token-based context 📑.

Core Architecture & Real-Time Performance

Mellum is optimized for the 'Flow' state, prioritizing UI responsiveness and predictive accuracy.

• Sparse MoE Switching: Dynamically swaps specialized experts based on the file extension and project manifest, reducing VRAM usage by 65% compared to dense 7B models 🧠.
• Next-Edit Suggestion (NES): A proprietary predictive layer that anticipates the next logical structural change (e.g., adding a try-catch block or implementing an interface) before the user starts typing 📑.
• Hardware-Specific Kernels: Optimized for Apple M4/M5 Neural Engine and NVIDIA RTX 50-series Tensor Cores, achieving sub-100ms response times for multiline completions 📑.

⠠⠉⠗⠑⠁⠞⠑⠙⠀⠃⠽⠀⠠⠁⠊⠞⠕⠉⠕⠗⠑⠲⠉⠕⠍

Sovereignty & Compliance Layer

The model architecture is a key component of the 'Air-Gapped AI' strategy for highly regulated sectors.

• Zero-Egress Execution: Inference is strictly local; the model does not require an active internet connection to provide structural suggestions, and no user code is used for global re-training 📑.
• Encrypted Weight Storage: Model weights (Safetensors) are cryptographically bound to the IDE license, preventing extraction or unauthorized fine-tuning 🧠.

Evaluation Guidance

Technical architects should prioritize the following validation steps:

• Symbol Resolution Depth: Benchmark Mellum's ability to suggest private internal methods in complex monorepos to verify the efficiency of the PSI-graph grounding 📑.
• NPU/GPU Thermal Profiling: Measure power consumption during 1-hour coding sessions on target hardware (e.g., MacBook Pro M4) to assess the impact of the MoE switcher on battery life 🧠.
• CAR Metrics: Analyze the Completion Acceptance Rate (CAR) across different languages to determine where the model requires fallback to high-parameter cloud models 📑.