What it does

This repo wraps MLX and the ds4 engine in an Anthropic-compatible API server so claude CLI talks to a local model instead of Anthropic’s cloud. Swap one environment variable and you swap the brain — Gemma 4 31B, Qwen 3.5 122B, or DeepSeek V4 Flash. The author also sells pre-configured Mac minis and $19 agent packs, but the core stack is open-source MIT.

The interesting bit

The “abliterated” angle: they package and upload their own MLX builds to HuggingFace, stripping alignment fine-tuning so the models are more pliable for agentic coding loops. For DeepSeek V4 Flash, they bolted on Antirez’s ds4 engine (pure C + Metal) within a day of release — giving 1M-token context and persistent disk KV cache so your 25k system prompt prefills once, ever.

Key highlights

• Qwen 3.5 122B at 65 tok/s via MLX-native MoE (only ~10B active params per token)
• DeepSeek V4 Flash: 284B params, ~37B active, 1M context, ~32 tok/s via ds4
• Gemma 4 31B: ~15 tok/s, fits 32 GB RAM, daily driver for 64 GB Macs
• Drop-in claude replacement scripts: claude-ds4, Claude Local.command, etc.
• Voice mode and browser-remote access via companion repos (Ears+Mouth, Hands, Phone)

Caveats

• RAM requirements are steep: 96 GB for Qwen, 128 GB for DeepSeek V4 Flash, and the 2-bit GGUF still eats 81 GB disk plus cache
• “Abliterated” means intentionally stripped safety tuning; the README treats this as a feature, but it’s a liability if you’re not airgapped
• Benchmarks shown are self-reported with YouTube demos; no independent replication cited

Verdict

Grab this if you’re a lawyer, doctor, or contractor who actually needs airgap guarantees and already owns a maxed-out Mac Studio. Skip it if you’re on a base M3 MacBook or if your threat model doesn’t justify trading cloud audit logs for locally-stripped model alignment.