SLM360

The First AI That Thinks, Learns, and Remembers On-Device

SLM360 is a complete, privacy-first AI system comprising two purpose-built foundation models (SLM360 Nano, a 6.4M-parameter encoder, and SLM360 Base, a 125M-parameter decoder), a hybrid NLU pipeline, and an on-device continual learning framework. All implemented in pure Rust with zero external ML dependencies.

Specifications

Architecture

 1  Tier 1: Pattern Matching (<1ms) - Regex, exact match, contains, fuzzy
 2  Tier 2: MicroTransformer, 85K params (2-5ms) - BPE tokenizer, 1-layer transformer
 3  Tier 3: SLM360 Nano, 6.4M params (<5ms) - Full encoder with GQA + SwiGLU
 4  Tier 4: SLM360 Base, 125M params (<50ms/tok) - Causal decoder for generation

Features

• Hybrid classification: pattern matching (<1ms) + semantic embeddings (39ms) with confidence arbitration
• Multi-step reasoning engine with conditional execution, sequences, and automatic rollback in <100ms
• 5-tier SmartMemory: short-term, episodic, semantic, procedural, and meta-learning
• Predictive context engine: anticipates user needs from topic transitions, time patterns, and entity preferences
• Cross-platform deployment: Linux, macOS, Windows, WebAssembly, with iOS and Android planned
• 100% on-device processing. Privacy-first by architecture, not policy
• 50MB total footprint: ONNX model (32MB), pattern engine (8MB), reasoning (4MB), SmartMemory (2MB), cache (3MB), runtime (1MB)
• 196 tests passing with comprehensive coverage across NLU, reasoning, memory, WASM, and async

Benchmarks

Deployment Targets

• >Native (ARM/x86) with SIMD auto-detection
• >WebAssembly (~300KB gzipped) for browser deployment
• >Android via JNI bindings
• >iOS via FFI bindings
• >Minimal mode (~50KB) for MCU deployment