Breaking

Open-source 'mac-code' runs 35B-class AI agents locally on a 16GB Mac mini M4

June 28, 2026 at 16:11 EDT

AI Agents
Open Source
Infra & Chips

A new open-source project called "mac-code" lets users run a 35B-parameter-class AI coding agent entirely locally and for free on a Mac mini M4 with 16GB of memory, using an Apple Silicon SSD-streaming technique dubbed "flash-paging." The GitHub repository by developer walter-grace aims to run Claude Code-style AI agents on consumer Macs without cloud APIs or expensive GPU servers.

Open Source · mac code

A 35B AI Coding Agent, Running Locally on a $600 Mac mini

A new open-source project streams an oversized model's weights from the SSD on demand — letting Qwen3.5-35B-A3B hit 30 tokens/sec on a 16GB Mac mini M4, using a technique it calls "flash-paging."

30 tok/s

Qwen3.5-35B-A3B (IQ2_M, recommended)

~5.5 GB

RAM used while keeping 4-bit quality

16 GB

Mac mini M4 RAM — runs a 35B model

Speed by mode — same model, same Mac

Tokens per second. The MoE + low-bit combo is the realistic choice; full-quality streaming collapses.

5.4

0.18

IQ2_M (MoE)
10.6 GB

Q4 Flash Stream
19.5 GB

27B Flash Stream
16.1 GB

How "flash-paging" fits a big model in small RAM

PINNED IN RAM

Attention, embeddings & KV cache stay resident.

→

STREAMED FROM SSD

FFN weights loaded on demand via direct I/O.

→

MoE: EXPERTS ONLY

Only the experts activated per token are fetched.

Received favorably

A 35B-class model on a ~$600 Mac mini M4 16GB, plus the mmap-based paging trick — set up entirely with brew and Python scripts.

Noted constraints

Flash Streaming is extremely slow (0.18 tok/s), so MoE + IQ2_M is the practical pick. Relies on mmap and needs a tool to pre-split the model.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.