Community shrinks MiniMax M3.1 Flash from 252GB to 93GB in 10 days, earning a thank-you from the lab
Dependent-Bid1391 · reddit · 2026-11-02
MiniMax publicly thanked community builders for quantizing the 428B-parameter M3.1 Flash (originally 252.4GB) within 10 days of release. Notable versions: PeasantSmith's Maya-XS at 93GB runs on a 128GB PC with 88.8% top-1 agreement on tool calls; ViC305's SAGE-EXL3 at 108.64GiB runs on one DGX Spark with 96.89% top-1 and 24-30 tok/s decode at 48K context; Unsloth's UD-Q2KXL at 143.9GB with 91.5% weighted top-1; superalesha ran Unsloth 2-bit on 4x 3090s with his own engine at 23-31 tok/s decode (needed 512GB RAM). Also oMLX for Apple Silicon, AMD Quark MXFP4 for MI350/MI355, and more. The model quantizes well because quantization calibration was baked into training, with sparse attention in every layer and only 23B of 428B params active per token. It's a coding/agent model with image/video input and up to 1M context. Numbers use different eval sets and aren't directly comparable.
More from Infra
- Gary Marcus flags CoreWeave's 57.88% six-month stock collapse ahead of Nov 10 earnings — GaryMarcus · 2026-11-03
- Pinokio ships "Remote AI" update for controlling local AI from anywhere — cocktailpeanut · 2026-11-03
- Pinokio Update Brings "Remote AI": Control Local AI Apps From Anywhere — cocktailpeanut · 2026-11-03
- Pinokio launches Remote: bring locally installed AI apps anywhere — cocktailpeanut · 2026-11-03
- Pinokio Remote Launches: Access Your Local AI Apps From Any Device — cocktailpeanut · 2026-11-03
- Wall Street's new narrative: hyperscalers turn cash flow positive in late 2027 — inductionheads · 2026-11-03