MiniMax M3.1 Flash: How a 428B MoE Model Cuts Agentic AI Costs by 70% Without Sacrificing Performance

MiniMax released M3.1 Flash on October 19, 2026, a 428-billion-parameter Mixture-of-Experts model that activates only 23 billion parameters per token, runs at over 200 tokens per second, and costs $0.15 per million input tokens — roughly 70% cheaper than its predecessor. The model pairs MiniMax Sparse Attention, which replaces the first three full-attention layers with sparse alternatives to support a practical 1-million-token context, with a new speculative decoding scheme called DSpark that ships with a dedicated draft head. Quantization-aware training, with routed experts at NVFP4 and the sparse-attention indexer and main branch at Q8KV4, lets the 428B model fit on NVIDIA Blackwell serving infrastructure.

MiniMax released M3.1 Flash https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/ on October 19, 2026, and the numbers are hard to ignore: a 428-billion-parameter Mixture-of-Experts model that activates only 23 billion parameters per token, runs at over 200 tokens per second, and costs $0.15 per million input tokens — roughly 70% cheaper than its predecessor. But the more interesting story is in the engineering choices that make those numbers possible. Mixture-of-Experts architectures have become the dominant approach for scaling language models without proportionally scaling inference cost. M3.1 Flash follows this pattern: the full model has 428 billion parameters spread across many expert sub-networks, but for any given token, only a small subset of those experts — totaling about 23 billion parameters — are activated by the router. This means the model carries the representational capacity of a very large dense network while paying the inference cost of a much smaller one. The 23B active parameter count puts it in the same compute class as a mid-sized dense model, but with access to the specialized knowledge encoded across the full expert pool. For agentic workloads that involve long multi-step tasks, this matters: you get deep reasoning without the per-token cost of a 400B+ dense model. Attention is the other major cost driver in transformer inference, especially at long context lengths. M3.1 Flash introduces MiniMax Sparse Attention MSA https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash , which replaces the first three full-attention layers with sparse alternatives. Full attention scales quadratically with sequence length — a 1-million-token context window would be computationally prohibitive without some form of sparsification. MSA selectively attends to a subset of positions rather than the full sequence, reducing the compute and memory footprint at long contexts while preserving the model's ability to track relevant information across the window. The first three layers handle early feature extraction, where full attention is most expensive and where sparse patterns tend to be sufficient. Later layers, which handle more complex reasoning, retain full attention where it matters most. The result is a model that genuinely supports a 1-million-token context window https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/ in practice, not just on paper. Speculative decoding is a well-established technique for accelerating autoregressive generation: a small "draft" model proposes several tokens at once, and the main model verifies them in parallel. If the draft is good enough, you get multiple tokens per forward pass of the large model. M3.1 Flash replaces the previous EAGLE-like Multi-Token Prediction MTP approach with DSpark https://build.nvidia.com/minimaxai/minimax-m3.1-flash/modelcard , a new speculative decoding scheme that ships with a dedicated draft head included in the model release. The draft head is purpose-built for M3.1 Flash's architecture, which means it can propose tokens that align well with the main model's distribution — reducing the rejection rate and improving the effective throughput gain. The practical outcome: MiniMax reports median output speeds of 170+ tokens per second, with peaks exceeding 200 TPS on compatible serving infrastructure NVIDIA Blackwell architectures — B200, B300, GB200, GB300 . For agentic applications where the model is generating long tool-call sequences or code blocks, this throughput difference is meaningful. To fit a 428B model into practical serving infrastructure, M3.1 Flash uses quantization-aware training QAT rather than post-training quantization. The routed MoE experts are quantized to NVFP4 W4A4 — 4-bit weights and 4-bit activations , while the sparse-attention indexer and main branch use Q8KV4. QAT matters here because it bakes the quantization error into the training process, allowing the model to compensate during learning rather than absorbing the accuracy hit after the fact. Post-training quantization at 4-bit precision can degrade performance noticeably on reasoning tasks; QAT narrows that gap significantly. The result is a model that achieves near-full-precision accuracy at a fraction of the memory footprint. One of the more practical features for developers is M3.1 Flash's configurable reasoning effort. The model supports five levels — low , medium , high , xhigh , and max — set via the chat template. This lets you trade off latency and token cost against reasoning depth depending on the task. A simple retrieval or classification task can run at low effort; a complex multi-step coding or cybersecurity task can run at max . Notably, the model enforces reasoning by default — attempting to disable it via API parameters returns an error. The design philosophy is that intermediate logic chains are always allocated, and the effort level controls how much deliberation happens, not whether it happens at all. MiniMax published results across a range of agentic benchmarks: The cybersecurity and coding numbers are strong. The Agents' Last Exam score of 29.6 reflects the genuine difficulty of that benchmark — it's designed to be hard for current models — rather than a model weakness. For comparison, the OSWorld V2 https://shattered.io/minimax-m3-1-flash-launch-cost-cut-2026/ score of 61.7 puts it among the top performers on computer-use tasks, which require the model to navigate real desktop environments across applications. The model accepts text, images, and video natively, making it multimodal out of the box. Inference is supported on vLLM, SGLang, and TokenSpeed runtimes. The 70% cost reduction relative to M3 is the headline, but the more durable point is the architectural coherence: MSA, DSpark, and QAT are not independent optimizations bolted on after the fact. They're designed together to make a 428B model serve at the cost and latency of a much smaller one. For teams building agentic pipelines — coding assistants, computer-use agents, financial automation — M3.1 Flash offers a practical option at the frontier of capability without the infrastructure cost of a dense model at comparable scale. The 1M-token context window and tunable reasoning effort give developers meaningful control over the cost-quality tradeoff at runtime. The model weights are available on Hugging Face https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash , and the API is live via MiniMax's platform at $0.15/$0.60 per million input/output tokens. For a detailed comparison against competing models, Kie.ai's benchmark breakdown https://kie.ai/blog/minimax-m3-1-flash-vs-deepseek-v4-1 is a useful reference. The NVIDIA model card https://build.nvidia.com/minimaxai/minimax-m3.1-flash/modelcard covers the full serving configuration details for Blackwell hardware.