MiniMax M3.1 Flash Review: 428B MoE at $0.15/Million Tokens
MiniMax's open-weight M3.1 Flash pairs a 428B/23B MoE with $0.15 per million input tokens, 200+ TPS output speed and vendor-reported agentic scores near the frontier tier.
The Cheapest Frontier-Adjacent Model of the Season
MiniMax officially rolled out M3.1 Flash on October 19, 2026, and the pricing alone changes the arithmetic for agent builders. The official blog states flat rates of US$0.15 per million input tokens, US$0.60 per million output tokens and US$0.015 per million cached input tokens (source: MiniMax Research blog, October 19, 2026). For comparison, the preceding M3 was listed at US$0.30 in and US$1.20 out per million tokens, which the same coverage describes as half of M3's per-token price (source: AICoder release coverage, October 2026).
There is also an origin story worth knowing. MiniMax confirmed that Space Bunny, an anonymous model that led usage rankings on OpenRouter and elsewhere, was an early version of M3.1 Flash, having processed more than 200 trillion tokens while operating anonymously (source: MiniMax Research blog, October 19, 2026). A model that ran anonymously at that volume under conditions the vendor does not control is a meaningful stress test, even though it is not a benchmark.
Architecture and Open Weights
The weights opened on Hugging Face on October 23, 2026 as MiniMaxAI/MiniMax-M3.1-Flash, ungated and distributed under the MiniMax Community license rather than MIT or Apache — a detail worth reading carefully before any commercial deployment (source: AICoder release coverage, October 2026). The checkpoint is MXFP8-quantised and is described as a native multimodal mixture-of-experts model with roughly 428 billion total parameters and about 23 billion active per token, with MiniMax Sparse Attention running throughout and a DSpark speculative-decoding draft head included (source: ModelScope model card for MiniMax-M3.1-Flash).
The context window is one million tokens, and the model accepts text, image and video input. Thinking cannot be switched off, and the effort parameter exposes five levels from low to medium, high, xhigh and max, with max as the default (source: AICoder release coverage, October 2026). That combination is unusual: a model where reasoning depth is a dial and non-negotiable, rather than a toggle.
The Vendor Numbers, Stated Plainly
Every figure below is vendor-reported, and no independent re-test was available at the time of writing — that caveat belongs at the front of any evaluation, not in a footnote.
On software engineering, M3.1 Flash scores 73.7% on DeepSWE v1.1, which the coverage places next to Opus 5 at 74.0% (source: AICoder release coverage, October 2026). On Automation-Bench it records 62.2%, and on BabyVision with tools 91.0%. The weakest reported result is Terminal Bench 4.0 at 38.1%, well behind Opus 5 at 51.8% — a meaningful gap for anyone betting on long-horizon terminal work.
On computer use, the official blog reports 61.7 points on OSWorld V2 and 29.6 points on Agents' Last Exam, describing the model as ranking first among Chinese models on both (source: MiniMax Research blog, October 19, 2026). On enterprise workflows the blog reports 76.2 on BankerToolBench and describes the model as ranking among the top performers on AutomationBench-AA. In security, the figure is 82% on CyberGym, with in-depth evaluations covering code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.
Throughput is the quiet headline. After more aggressive sparsity optimisation, the blog reports output speeds reaching 200+ tokens per second in everyday use, with median (P50) speeds consistently above 170 TPS (source: MiniMax Research blog, October 19, 2026). The blog also claims overall usage cost 70 percent lower than the predecessor for agentic tasks, and online inference nearly three times faster; a separate coverage notes MiniMax stating the model costs one third of M3 and runs 3x faster, while flagging that the vendor does not explain how the one-third figure and the 70 percent figure reconcile (source: AICoder release coverage, October 2026).
Ecosystem and Practical Notes
Day-zero serving was not limited to the vendor's own API. Together AI runs the model fully managed; MiniMax is partnering with Inferact to feed production-inference improvements back into vLLM; and LightSeek's TokenSpeed added support on launch day (source: AICoder release coverage, October 2026). MiniMax Code is available as a desktop download, M Plan as a subscription, and API access through platform.minimax.io. M2.5 and M2.7 were closed to new users on October 19 and are scheduled to retire on November 18, with M Plan and Token Plan requests routing to M3.1 Flash automatically.
Who Should Buy It
The case for M3.1 Flash is narrow and strong: high-volume agent loops, tool-calling pipelines, and code tasks where per-token cost dominates and a human is supervising the output. At US$0.15 in and US$0.60 out, the margin for error is far wider than with frontier pricing, and the 1M context plus 200+ TPS makes long-context loops practical rather than theoretical.
The case against is equally clear. Terminal Bench 4.0 at 38.1% says this is not yet a model for unattended long-horizon terminal autonomy, and all benchmark numbers remain vendor-reported with no independent re-test available. The MiniMax Community license is the quiet risk for any commercial deployment: it is not a permissive open-source license, and it should be read before any business decision, not after.
On balance, for agentic workloads at scale, the pricing-speed combination is genuinely hard to argue with. For frontier-quality autonomous work, the published evidence does not yet support the switch. Read the license, run your own evals, and let the numbers decide.