We're excited to be an official Day 0 provider for MiniMax M3.1 Flash. Built for coding, agents, and complex workflows, M3.1 Flash brings strong agentic performance at a competitive cost, and it's available today on our serverless API.

Try it on Inference or sign up free to get started.

What is MiniMax M3.1 Flash?

MiniMax M3.1 Flash is a native multimodal Mixture-of-Experts model. It has roughly 428B total parameters and activates 23B per token, which keeps latency and cost low while preserving deep reasoning ability.

It builds on MiniMax-M3 with three changes aimed at efficiency:

Sparse attention throughout. MiniMax Sparse Attention now runs in every layer, including the first three layers that used full attention in M3.
Quantization-aware training. The routed experts and attention paths are quantized during training, which reduces the memory footprint and improves serving efficiency.
DSpark speculative decoding. A new draft head replaces the earlier MTP approach to speed up generation.

On Inference, M3.1 Flash supports a 1M-token context window, image input, function calling, structured outputs, and automatic prompt caching. Reasoning effort is selectable per request: low, medium, high, xhigh, or max.

Benchmarks

MiniMax reports strong results on agentic, coding, and computer-use benchmarks. M3.1 Flash leads its comparison set on Automation-Bench (62.8) and LOCA-Bench 256K (60.3), and it ties Opus 5 on BankerToolBench (76.2).

We tried it out

Benchmarks only tell part of the story, so we gave M3.1 Flash a few open-ended prompts: an interactive double-pendulum lab, a 3D datacenter technician game, and of course, the classic. You can watch the first two in our launch thread on X.