MiniMax M3.1 Flash (CanopyWave)

MiniMax officially launched M3.1 Flash on October 19, 2026, identifying the previously anonymous Space Bunny model as an early version. The company targeted complex agentic work, coding, computer use, and cybersecurity, and said aggressive sparsity improved online generation speed and reduced operating expenses by 70%. MiniMax reported latency falling from 1.95 to 1.22 seconds and throughput rising from 70 to 100 tokens per second. It also reported 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, and 29.6 on Agents’ Last Exam, while standard API prices were set at $0.15 per million input tokens and $0.60 per million output tokens.

BenchmarkList’s MiniMax M3.1 Flash tracker lists 22 evaluations, including Artificial Analysis results for AutomationBench-AA at 72.5% partial score, AA-Briefcase at 1,473 Elo, and PostTrainBench at 39.3% average. The page separately identifies the 29.6% Agents’ Last Exam pass rate as self-reported. The tracker places MiniMax M3.1 Flash fourth among 32 models on AutomationBench-AA, 31st among 16 on AA-Briefcase, and eighth among 35 on PostTrainBench. It also lists MiniMax’s direct pricing at $0.15 per million input tokens and $0.60 per million output tokens.

Updated October 27, 2026, this release tracker describes MiniMax M3.1 Flash as keeping the MiniMax-M3 mixture-of-experts design, at roughly 428B total and 23B active parameters, while adding sparse attention throughout, quantization-aware training, and DSpark speculative decoding, with a 1M token context window. The tracker lists the model as generally available and the family as MiniMax M3, following MiniMax-M3. MiniMax says online inference is nearly three times faster at 200+ tokens per second in everyday use, with agentic task cost 70% below the prior generation, and reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, and 82% on CyberGym. Access is via the MiniMax API, the M Plan subscription, the MiniMax Code desktop app, and a Hugging Face repository under the minimax-community license, with pricing at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. Results are self-reported, and usable context and throughput depend on GPU memory and serving setup.

MiniMax released M3.1 Flash on October 19, 2026, a model aimed at multi-step agent workflows, software engineering, computer use, and security assistance. The launch wire describes enterprise use across sales, marketing, operations, finance, and HR SaaS tools, plus long development workflows spanning code understanding through implementation and validation. MiniMax reports nearly three times faster online inference, median output above 170 tokens per second, and roughly 70% lower agent-task cost than the previous generation, with a reported 73.7% on DeepSWE v1.1. The model is available through the MiniMax API, the M Plan subscription, and MiniMax Code, at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. These performance and cost figures are company-reported, and weights were promised for a later release.

An unofficial InferenceX breakdown derived from the published checkpoint config describes MiniMax M3.1 Flash as a 60-layer MoE backbone with hidden size 6,144 and vocabulary 200,064, where the first three layers use dense feed-forward blocks and the remaining 57 use MoE. Attention is MiniMax Sparse Attention in all 60 layers, unlike M3, whose first three attention layers were full GQA. It reports 128 routed experts with top-4 selection plus one shared expert, GQA with 64 query and 4 KV heads, per-head QK normalization, partial RoPE with factor 0.5, swigluoai activation at alpha 1.702, a 1,048,576 config context limit, NVFP4 experts with MXFP8 dense weights, and a built-in DSpark draft head. The page states it is unaffiliated with SemiAnalysis and its data may be modified.

NVIDIA's NIM model card for minimax-m3.1-flash states the model is developed by MiniMax, not NVIDIA, and is built for coding, agentic workflows, computer use and cybersecurity with text and image input, text output, and a context window up to 1M tokens. It cites roughly 428B total parameters with 23B active per token. The card adds deployment details absent from the launch post: quantization-aware training to NVFP4, MiniMax Sparse Attention across the model, DSpark speculative decoding, and five configurable reasoning effort levels (low, medium, high, xhigh, max). Release is listed as 10/22/2026 on build.nvidia.com and Hugging Face, with the base model under the MiniMax Community License and research-only trial terms.

MiniMax's official Hugging Face card describes MiniMax-M3.1-Flash as a native multimodal mixture-of-experts model for coding, agentic workflows, computer use and cybersecurity, with roughly 428B total parameters and 23B activated per token, text/image/video inputs, and a context window up to one million tokens. It builds on MiniMax-M3 using sparse attention throughout, quantization-aware training, and DSpark speculative decoding. The model card lists agentic scores of 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, 29.6 on Agents' Last Exam, and 82.2 on CyberGym, alongside a comparison table against DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 and Opus 5, where it trails those rivals on several suites.

MiniMax officially rolled out MiniMax M3.1 Flash, which it calls a next-generation Flash model delivering a generational leap in agentic tasks, coding, computer use and cybersecurity, while making online inference nearly three times faster and cutting usage costs by 70%. The company revealed that Space Bunny, the anonymous model that processed more than 200 trillion tokens, was an early version of M3.1 Flash. MiniMax reports M3.1 Flash ranks among the top performers on AutomationBench-AA for real enterprise SaaS workflows, scores 76.2 on BankerToolBench for multi-step financial tool calls, and reaches 73.7% on DeepSWE v1.1 for independent software development tasks. On computer use it posts 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam, ranking first among Chinese models on both.