MiniMax thanks community quantizers running its 428B M3.1 Flash locally at 143.9GB
MiniMax_AI · x · 2026-11-02
MiniMax publicly thanked community builders for low-bit quantizations of its 428B-parameter M3.1 Flash, enabling local runs on more hardware. Highlights: UnslothAI's 2-bit dynamic GGUF at 143.9 GB with calibration and fidelity tests across coding and tool use; ViC305's SAGE-EXL3 token-level fidelity measurements and single DGX-Spark recipe; Project Maya expert offloading; oQ quants via oMLX on Apple Silicon; AMD Quark MXFP4 for Instinct; and custom inference combining quantized weights with expert caching on 4x RTX 3090. MiniMax notes making low-bit savings useful requires careful calibration, efficient kernels, and hardware-aware expert placement across GPU and system RAM.
More from Infra
- FT chart: Big Tech's surging CapEx is a rational bet, not reckless spending — luisdans · 2026-11-02
- Liquid AI's 2B d1 decision models beat same-size general models on classification benchmarks — JosephJacks_ · 2026-11-02
- Intelligence got 100x cheaper, so usage exploded 500x: Jevons Paradox in action — luisdans · 2026-11-02
- OpenAI may rebrand its data centers as 'AI compute campuses' — TechCrunch AI · 2026-11-02
- $300B Louisiana project pipeline squeezes steel supply, H-beam lead times hit 9 months — MatthewChang · 2026-11-02
- Benchmark: Speedy inference engine crushes Avalon — 1.6x faster reads on Qwen3 long context — MotorNetwork380 · 2026-11-02