JustFit squeezes 200K-token context onto a 24GB MacBook
The MLX runtime reaches 213K tokens on an M4 Pro without quantizing weights — one of several fresh tricks pushing big-model inference onto your own laptop.
Copy markdown
213K tokens on a laptop, no weight-quant crutch
On a 24GB M4 Pro, JustFit serves Qwen 3.8-27B (MXFP4) at 212,992 tokens in a single request — 6.9x the 30,720-token baseline — holding ~19 tok/s and a ~16GB peak, and it cleared 29 of 30 AIME 2026 problems. The gains come from compressed-KV execution plus swapping model components in and out of RAM, so they stack on top of quantization instead of competing with it.
The fine print: it's a paper, not a package
There's no public repo yet, so you can't install JustFit today — it's a systems paper, not a release. But the three moving parts (KVExec, PhaseSwap, StateTrans) are spelled out clearly enough to reimplement in an MLX pipeline, and because they manage execution state rather than weights, they layer onto whatever quantization you already run.
Build this weekend: losslessly shrink BitNet weights
Want something runnable now? The open bitnet-t5b repo re-encodes BitNet b1.58 weights to 1.60 bits/weight (down from 2.0) with table-free AVX2 decoding and bit-identical output — one model drops 1.10 to 1.00 GiB and generates ~1.13x faster at the same 96.82 perplexity. The underlying BITCOS paper found zeros make up as much as 51.5% of ternary weights, letting a distribution-aware layout hit 1.485 bits and break the 1.585-bit 'floor'; `make test` runs 177K correctness checks.
Elsewhere: LimiX-2 rethinks tabular ML from the ground up
A new structured-data foundation model, LimiX-2, claims to top both dataset-specific models and other tabular foundation models on TabArena, TALENT and BCCO by learning the joint p(x,y|D) instead of the usual p(y|x). Weights aren't out yet, but if you build on tabular data it's the paper to watch after this month's TabPFN-3.5 run.