A paper posted to arXiv on 15 September reports running a 27-billion-parameter model with a 200,000-token context on a laptop. The system, called JustFit, is built on Apple’s MLX framework and combines compressed key-value execution, swapping model components in and out of memory, and preserving state between requests.
The reported numbers are specific. On a 24 GiB M4 Pro MacBook running Qwen3.8-27B in MXFP4, the author reports single-request context rising from 30,720 to 212,992 tokens — a factor of just under seven — at 19.11 tokens per second, with a peak memory footprint of 16,374 MiB. The paper also claims 29 of 30 AIME 2026 problems answered correctly across a run of roughly 197,000 input and 16,000 output tokens.
Treat all of that as self-reported. This is a single-author preprint, it has not been peer-reviewed, and the benchmark figures come from the same person who built the system; nobody outside has reproduced them. The paper is arXiv:2609.17475.
Why it matters: if long-context inference genuinely fits on consumer hardware, the assumption that serious reasoning workloads must run in someone else’s data centre gets weaker. That is a large claim resting on one unreplicated paper, which is why it is worth watching rather than believing.