利用苹果“LLM in a Flash”在本地运行Qwen 397B的研究

18th March 2026 - Link Blog

Autoresearching Apple's "LLM in a Flash" to run Qwen 397B locally. Here's a fascinating piece of research by Dan Woods, who managed to get a custom version of Qwen3.5-397B-A17B running at 5.5+ tokens/second on a 48GB MacBook Pro M3 Max despite that model taking up 209GB (120GB quantized) on disk.

Qwen3.5-397B-A17B is a Mixture-of-Experts (MoE) model, which means that each token only needs to run against a subset of the overall model weights. These expert weights can be streamed into memory from SSD, saving them from all needing to be held in RAM at the same time.

Dan used techniques described in Apple's 2023 paper LLM in a flash: Efficient Large Language Model Inference with Limited Memory:

This paper tackles the challenge of efficiently running LLMs that exceed the available DRAM capacity by storing the model parameters in flash memory, but bringing them on demand to DRAM. Our method involves constructing an inference cost model that takes into account the characteristics of flash memory, guiding us to optimize in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks.

He fed the paper to Claude Code and used a variant of Andrej Karpathy's autoresearch pattern to have Claude run 90 experiments and produce MLX Objective-C and Metal code that ran the model as efficiently as possible.

danveloper/flash-moe has the resulting code plus a PDF paper mostly written by Claude Opus 4.6 describing the experiment in full.

The final model has the experts quantized to 2-bit, but the non-expert parts of the model such as the embedding table and routing matrices are kept at their original precision, adding up to 5.5GB which stays resident in memory while the model is running.

Qwen 3.5 usually runs 10 experts per token, but this setup dropped that to 4 while claiming that the biggest quality drop-off occurred at 3.

It's not clear to me how much the quality of the model results are affected. Claude claimed that "Output quality at 2-bit is indistinguishable from 4-bit for these evaluations", but the description of the evaluations it ran is quite thin.

Update: Dan's latest version upgrades to 4-bit quantization of the experts (209GB on disk, 4.36 tokens/second) after finding that the 2-bit version broke tool calling while 4-bit handles that well.