MiniMax-M3.1-Flash joins GLM and DeepSeek as Maya's third model, on an engine of its own, with MiniMax's own
settings and fitted to each PC from the first release.
- MiniMax-M3.1-Flash on its own engine (#477, @needmorevram):
./setup.sh --model MiniMax-M3.1-Flashdownloads
Mia-AiLab's EXL3 3.5 bpw quant (197 GB, every file checked) and runs it onmaya-mm3, the engine @needmorevram
built for this model: tensor parallel over groups of GPUs, the DSpark drafter (about 1.5x decode in his runs),
pictures, a streamed K/V cache, and the experts VRAM cannot hold computed on the CPU from RAM. On an RTX 3090 with
eight RTX 5060 Ti he measured 38.2 tokens/s writing code. - Tensor parallel and tool calls, ready at launch (#490, @needmorevram): GPU groups read long prompts
together, the INT8 K/V cache keeps its quality, and MiniMax tool calls keep string values as strings. - Windows and V100-class GPUs: MiniMax runs wherever DeepSeek's EXL3 does: NVIDIA from V100 up, on Linux and
Windows (Windows is new: tell us how it goes), with the V100, T4 and RTX 20 tensor cores in use. - A compact K/V cache by default: INT8 takes half of BF16's memory, so a longer context or more experts fit;
fp8,fp4and mixes are options (--kv-cache). - Pictures on, never parked in VRAM: the vision tower stays in RAM or on the SSD and comes to the GPU only while
a picture is read, so the VRAM goes to the model. - Fitted to each PC: the setup recommends the context (up to the model's 1M) by the PC's memory, the engine plans
VRAM and RAM at start and steps down with a note instead of failing, the DSpark drafter stays only where it measures
faster, and the CPU lane uses the PC's own cores. - MiniMax's own settings: the model card's sampling (temperature 1.0, top-p 0.95), MiniMax's official prompt
format, and its thinking levels in the chat: Auto (the default, the model decides whether to think), Low to Max,
and Off. - Fix: an empty answer after an engine error (DeepSeek and MiniMax): the request after a failed one (an
unreadable picture, for example) now gets its own answer. - New setting:
"reasoning_effort": "auto"(any model) asks for the model's own default level.
MiniMax runs on NVIDIA V100 / RTX 20 and newer, Linux and Windows, with about 250 GB of VRAM and RAM together. Its license,
the MiniMax Community License (LICENSE.minimax-m31), allows non-commercial use freely; commercial use has conditions.
Update: the dashboard's About > Update, or git pull, then ./setup.sh (Windows: START-MAYA.bat) - it recompiles only what changed; the model is not downloaded again.