JustFit: 200K-Token LLM Serving on a 24 GiB Laptop with Just-in-Time State Management
Paper • 2609.17475 • Published
How to use billchen42/JustFit-Qwen3.8-27B-components with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir JustFit-Qwen3.8-27B-components billchen42/JustFit-Qwen3.8-27B-components
Published checkpoint-side components for the JustFit reproduction kit and paper.
This repository does not duplicate the target model. Download the pinned
mlx-community/Qwen3.8-27B-mxfp4 target separately.
| Path | Bytes | SHA-256 |
|---|---|---|
vision-bf16.safetensors |
921,492,862 | b27ad7963b3752b5cab9d225f9c0b033d0d59066e7ff357dac17d690a6ff7994 |
mtp/model.safetensors |
225,662,258 | 9b1d99f402f98940e08891d04815271e3fa94d6bc762f042fd311d3aa0d8bd1d |
mtp/config.json |
3,777 | e0d0d5dc68f59559940e1b2dafc40a34700b9ae7ff963de268225a54550af563 |
vision-bf16.safetensors is the exact BF16 backing used by the production
PhaseSwap vision adapter. Its 333 tensors were matched tensor-for-tensor to
the pinned public MXFP4 checkpoint.mtp/ is the standalone block-size-3 MXFP4 MTP drafter. Its weight file is
byte-identical to the measured artifact.hf download billchen42/JustFit-Qwen3.8-27B-components \
--revision 13e0462fa911a1bcbdccbaa759120500c0f90582 \
--local-dir ./justfit-components
For the full launcher, capacity client, pinned target revision, limitations, and result definitions, use the linked reproduction kit rather than treating these files as a standalone model.
Quantized
Base model
Qwen/Qwen3.8-27B