walter-grace

Mac Code — AI skill for Claude Code

AI community

mac code — Claude Code, but it runs on your Mac for free.

How to install Mac Code

This entry records only its repository, not the path inside it, so there is no exact command to give. Open walter-grace/mac-code and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Mac Code does

mac code — Claude Code, but it runs on your Mac for free. 35B AI agent at 30 tok/s via Apple Silicon flash-paging. $0/month.

Alternatives in AI

System Prompts Leaks — Extracted system prompts from ChatGPT (GPT-5.4, GPT-5.3, Codex), Claude (Opus 4.6, Sonnet 4.6, Claude Code), G 38.6k ★
Optimization — ai-research-skills AWQ, GPTQ, GGUF, Flash Attention, bitsandbytes 5.4k ★
CodeIsland — Real-time AI coding agent status panel in your MacBook notch — live status, approvals & replies for 13 AI tool 2.3k ★

README

mac code

**Run models that don't fit in RAM on your Mac. $0/month.**

Can I run this on my Mac?

Your Mac	RAM	What you can run	Speed
Any Mac	8 GB	Qwen3.5-9B (Q4_K_M, 5.3 GB), 4K context	16-20 tok/s
Any Mac	16 GB	Qwen3.5-9B (Q4_K_M, 5.3 GB), 64K context	16-20 tok/s
Mac mini M4	16 GB	Qwen3.5-35B-A3B (IQ2_M, 10.6 GB)	30 tok/s
Mac mini M4	16 GB	Qwen3-30B-A3B Q4 (17.2 GB) via Expert Sniper	4.3 tok/s
Mac mini M4	16 GB	Qwen3.5-35B-A3B Q4 (19.5 GB) via Expert Sniper	5.4 tok/s
Mac mini M4	16 GB	Qwen3.5-35B-A3B Q4_K_M (22 GB) via Flash Streaming	1.54 tok/s
Mac mini M4	16 GB	Qwen3.5-27B (16.1 GB) via Flash Streaming	0.18 tok/s
Mac mini M4 Pro	48 GB	35B at full Q4 in RAM	30+ tok/s

**"I wanted to run the Qwen 27B on my M2 16GB but failed. That's not possible, right?"**

It is possible. We stream FFN weights from SSD — only 5.5 GB stays in RAM. The output is coherent, full 4-bit quality. It's slow (0.18 tok/s on a Mac mini M4) but the method works on any 16 GB Apple Silicon Mac. No 2-bit compression, no mmap thrashing, no swap death. [See how it works.](#how-flash-streaming-works)

Quick Start

35B Agent (recommended — 30 tok/s on 16 GB)

The fastest option. Uses llama.cpp with a 2-bit quantization (IQ2_M) that fits entirely in RAM.

brew install llama.cpp
pip3 install rich ddgs --break-system-packages

# Download model (10.6 GB)
python3 -c "
from huggingface_hub import hf_hub_download
hf_hub_download('unsloth/Qwen3.5-35B-A3B-GGUF',
    'Qwen3.5-35B-A3B-UD-IQ2_M.gguf', local_dir='$HOME/models/')
"

# Start server + agent
llama-server \
    --model ~/models/Qwen3.5-35B-A3B-UD-IQ2_M.gguf \
    --port 8000 --host 127.0.0.1 \
    --flash-attn on --ctx-size 12288 \
    --cache-type-k q4_0 --cache-type-v q4_0 \
    --n-gpu-layers 99 --reasoning off -np 1 -t 4

python3 agent.