PheronAgent Under the Hood
A hardware-native autonomous AI agent for macOS. Built on Swift 6 UNO architecture, running entirely on Apple Silicon via local MLX inference.
Built for Extreme Performance
Pheron Agent is built with native Apple hardware components to achieve speeds unmatched by cloud-based alternatives.
Titan Engine
On-device MLX inference featuring wired memory pinning, 4-bit KV quantization, rotating cache (up to 131K context), and speculative decoding via custom draft models.
ANE Intent Classifier
Hardware-accelerated task routing executed directly on the Apple Neural Engine. Routes prompts to tools, weather, chat, or LLM fallback in milliseconds.
Three-Layer Memory
Structured memory layers: L1 Hot Cache (12 messages), L2 Daily Notes, and L3 DreamBank long-term summaries coupled with Metal-accelerated RAG via custom Metal kernels.
Energy Profiling
Monitored via `PheronEnergyDaemon` XPC helper utilizing `powermetrics` at 100ms intervals for exact, hardware-level Joule accounting per task execution.
Privacy Guard
Rule-based + local LLM PII (Personally Identifiable Information) detection before any external routing, executing PASS, DESENSITIZE, or BLOCK decisions.
SkillVault
Self-improving procedural memory. The agent writes and patches its own `.skill.md` tool scripts, while a background curator Actor consolidates skills across sessions.
Supported Models
All models run entirely on-device via MLX. No internet required. Tool calling and thinking mode support varies by architecture.
| Model | Architecture | Quantization | Min RAM | Speed (M4) | Tool Calling | Thinking Mode |
|---|---|---|---|---|---|---|
| Qwen3.5 4B Hybrid GatedDeltaNet + full-attention architecture. | Qwen 3.5 | 4-bit MLX | 6 GB UMA | ~80 tok/s | ✓ | ✓ |
| Qwen3.5 9B Default model for 16 GB devices. | Qwen 3.5 | 4-bit MLX | 10 GB UMA | ~50 tok/s | ✓ | ✓ |
| Qwen3.5 9B OptiQ Mixed-precision quantization for higher quality. | Qwen 3.5 | OptiQ 4-bit | 10 GB UMA | ~50 tok/s | ✓ | ✓ |
| Qwen3.5 27B Ideal for 24–32 GB Macs. | Qwen 3.5 | 4-bit MLX | 18 GB UMA | ~20 tok/s | ✓ | ✓ |
| Llama 3.2 1B Ultra-lightweight model. | Llama 3.2 | 4-bit MLX | 2 GB UMA | ~180 tok/s | ✓ | — |
| Llama 3.2 3B Default fallback for base M1/M2 chips. | Llama 3.2 | 4-bit MLX | 4 GB UMA | ~120 tok/s | ✓ | — |
| Llama 3.1 8B Standard Llama 3.1 8B instruct. | Llama 3.1 | 4-bit MLX | 8 GB UMA | ~35 tok/s | ✓ | — |
| Llama 3.3 70B Advanced reasoning for Workstation-class systems. | Llama 3.3 | 4-bit MLX | 48 GB UMA | ~10 tok/s | ✓ | — |
| Llama 4 Scout ⚠️ Experimental — Pending mlx-swift-lm llama4 support. | Llama 4 MoE | 4-bit MoE | 80 GB UMA | ~8 tok/s | ✓ | — |
| Llama 4 Maverick ⚠️ Experimental — Massive MoE model. | Llama 4 MoE | 4-bit MoE | 512 GB UMA | ~3 tok/s | ✓ | — |
| Gemma 3 1B Lightweight and fast Gemma model. | Gemma 3 | 4-bit MLX | 2 GB UMA | ~250 tok/s | — | — |
| Gemma 3 4B Balanced Gemma 3 performance. | Gemma 3 | 4-bit MLX | 6 GB UMA | ~85 tok/s | — | — |
| Gemma 3 12B High-capability Gemma for 16 GB RAM systems. | Gemma 3 | 4-bit MLX | 12 GB UMA | ~35 tok/s | — | — |
| Gemma 3 27B Gemma 3 model for complex tasks. | Gemma 3 | 4-bit MLX | 24 GB UMA | ~20 tok/s | — | — |
| Gemma 4 E4B Next-gen dense 4.5B model with tool calling. | Gemma 4 | 4-bit MLX | 8 GB UMA | ~60 tok/s | ✓ | ✓ |
| Gemma 4 26B ⚠️ Experimental — Pending MoE router fix. | Gemma 4 MoE | 4-bit MoE | 20 GB UMA | ~25 tok/s | ✓ | ✓ |
| Mistral 7B v0.3 Reliable Mistral 7B model. | Mistral | 4-bit MLX | 8 GB UMA | ~45 tok/s | ✓ | — |
| Mistral Nemo 12B Balanced model with large 128K context window. | Mistral | 4-bit MLX | 12 GB UMA | ~30 tok/s | ✓ | — |
| Mistral Small 24B Mistral Small for general-purpose tasks. | Mistral | 4-bit MLX | 16 GB UMA | ~18 tok/s | ✓ | — |
| Mistral Small 3.2 24B Mistral Small 3.2 with enhanced tool calling. | Mistral | 4-bit MLX | 16 GB UMA | ~18 tok/s | ✓ | — |
| Devstral Small 24B ⚠️ Experimental — Developer-focused model. | Mistral | 4-bit MLX | 24 GB UMA | ~17 tok/s | ✓ | — |
| Mistral Large 123B Large-scale Mistral Large model. | Mistral | 4-bit MLX | 128 GB UMA | ~6 tok/s | ✓ | — |
| Devstral 2 123B ⚠️ Experimental — Pending ministral3 architecture. | Mistral | 4-bit MLX | 128 GB UMA | ~6 tok/s | ✓ | — |
| Phi-4 Mini Microsoft reasoning model with native function calling. | Phi 4 | 4-bit MLX | 4 GB UMA | ~150 tok/s | ✓ | — |
| Phi-4 14B Phi-4 14B instruct model. | Phi 4 | 4-bit MLX | 12 GB UMA | ~35 tok/s | — | — |
| DeepSeek Coder V2 Lite ⚠️ Experimental — Pending deepseek_v2 architecture. | DeepSeek MoE | 4-bit MoE | 12 GB UMA | ~30 tok/s | — | — |
| DeepSeek V4 Flash ⚠️ Experimental — Pending deepseek_v4 architecture. | DeepSeek MoE | 4-bit MoE | 192 GB UMA | ~4 tok/s | ✓ | ✓ |
| Qwen2.5-VL 3B Vision-Language model requiring 24 GB+ UMA. | Qwen 2.5 VL | 4-bit MLX | 24 GB UMA | ~20 tok/s | — | — |
| Qwen3-VL 4B Vision-Language model requiring 32 GB+ UMA. | Qwen 3 VL | 4-bit MLX | 32 GB UMA | ~18 tok/s | — | — |
| Qwen2.5-VL 7B Vision-Language model requiring 48 GB+ UMA. | Qwen 2.5 VL | 4-bit MLX | 48 GB UMA | ~10 tok/s | — | — |
Performance by Chip
Pheron Agent automatically selects the best model for your hardware. All tiers require macOS 26.0+ and Apple Silicon.
- ✓ Titan Engine
- ✓ ANE Intent Classifier
- ✓ 62 Tools
- ✓ Quantized KV Cache
- ✓ Titan Engine
- ✓ ANE Intent Classifier
- ✓ 62 Tools
- ✓ Speculative Decoding
- ✓ Semantic Vision (24GB)
- ✓ Titan Engine
- ✓ ANE Intent Classifier
- ✓ 62 Tools
- ✓ Speculative Decoding
- ✓ 65K Context (8-bit KV)
- ✓ Titan Engine
- ✓ ANE Intent Classifier
- ✓ 62 Tools
- ✓ Speculative Decoding
- ✓ 131K Context (FP16 KV)
Speed figures are benchmarked on M4. M1/M2 devices run approximately 2–3× slower on equivalent models. Semantic VLM requires 24 GB+ unified memory.