MiniMax M3.1 Flash pairs 428B MoE with sparse attention to cut agentic inference costs

A 428B model with a mid-size compute bill

MiniMax released M3.1 Flash on October 19, 2026. According to a dev.to analysis of the release, it is a 428-billion-parameter Mixture-of-Experts (MoE) model that brings only about 23 billion parameters into play for each token, generates at a median of more than 170 tokens per second with peaks above 200, and is priced at $0.15 per million input tokens and $0.60 per million output tokens — around 70% cheaper than its predecessor, M3.

The MoE pattern is now the standard route to growing model capacity without growing inference cost in lockstep. In M3.1 Flash, the 428 billion parameters are spread across many expert sub-networks, and a router picks the small subset — totalling roughly 23 billion parameters — that processes any given token. The effect, as the dev.to write-up frames it, is access to knowledge encoded across the entire expert pool while paying a compute bill closer to a mid-sized dense model. For agentic workloads built on long multi-step chains, that trade is the whole point: frontier-level capability without per-token pricing that behaves like a 400B-class dense model.

Sparse attention for million-token contexts

The other major cost driver in transformer inference is attention, which scales quadratically with sequence length. M3.1 Flash introduces MiniMax Sparse Attention (MSA), which replaces the full-attention mechanism in the first three layers with sparse alternatives that attend to a subset of positions instead of the whole sequence. The dev.to analysis notes a rationale for which layers were changed: the earliest layers do feature extraction, where sparse patterns tend to be enough, while later layers responsible for harder reasoning keep full attention. The upshot is a 1-million-token context window that the write-up describes as genuinely usable rather than a spec-sheet figure.

A purpose-built draft head for faster decoding

Speculative decoding lets a small draft model propose several tokens at once, which the main model verifies in parallel. M3.1 Flash drops the earlier EAGLE-style multi-token prediction setup in favour of DSpark, a scheme that ships with a dedicated draft head tailored to the model's own architecture. Because the draft head's proposals align well with the main model's output distribution, fewer tokens are rejected and the throughput gain survives in practice. MiniMax reports median output speeds above 170 tokens per second, peaking beyond 200 TPS on NVIDIA Blackwell hardware (B200, B300, GB200, GB300) — a difference that matters when agents emit long tool-call sequences or code blocks.

Four-bit precision trained in, not bolted on

Fitting 428 billion parameters onto practical serving hardware required quantization-aware training (QAT) rather than post-training quantization. The routed MoE experts are quantized to NVFP4 (4-bit weights and 4-bit activations), while the sparse-attention indexer and the main branch use Q8KV4. QAT exposes the model to quantization effects during training so it can learn around them; plain post-training 4-bit quantization is known to hurt reasoning tasks. The result, per the analysis, is accuracy close to full precision at a fraction of the memory footprint.

Reasoning effort is a dial, not a switch

Developers can set reasoning effort to one of five levels — low, medium, high, xhigh or max — through the chat template. Reasoning cannot be disabled: the API returns an error if you try. The design intent is that the model always runs an intermediate reasoning pass, and the dial only changes how long that deliberation runs.

Benchmarks and availability

MiniMax's published scores include 82% on CyberGym (vulnerability discovery), 76.2 on BankerToolBench (financial tool use), 73.7% on DeepSWE v1.1 (software engineering), 61.7 on OSWorld V2 (computer use) and 29.6 on Agents' Last Exam. The dev.to analysis reads the cybersecurity and coding results as strong, places the OSWorld V2 score among the top performers on desktop-navigation tasks, and attributes the low Agents' Last Exam figure to that benchmark's difficulty rather than a model weakness. The model accepts text, images and video natively, runs on the vLLM, SGLang and TokenSpeed runtimes, and its weights are available on Hugging Face.

Why it matters

Agentic pipelines multiply per-token costs: multi-step tasks, long contexts and verbose tool-call output all compound on every request. M3.1 Flash's significance lies less in any single number than in the fact that its three core techniques — sparse attention, a matched speculative draft head and training-time quantization — were engineered as one system, so a 428B model can be served at the cost and latency of a far smaller one. For teams building coding assistants, computer-use agents or financial automation, that promises frontier capability without dense-scale infrastructure. One caveat: the performance, speed and cost figures all originate with MiniMax and are relayed through a single dev.to write-up, so independent verification is still pending.