MiniMax Launches M3.1 Flash on API: 428B MoE, 1M Context, Reasoning Always On
According to a MarkTechPost report dated October 19, 2026, MiniMax has launched M3.1 Flash on its API: 428B total parameters, roughly 23B activated, 1M token context, priced at $0.15 input and $0.60 output per million tokens. MiniMax said weights are expected on October 23, with the license not yet disclosed.
What changed
According to a MarkTechPost report dated October 19, 2026, MiniMax has launched MiniMax M3.1 Flash on its API (platform model name MiniMax-M3.1-Flash), positioned as a multimodal MoE model for coding, agentic office tasks and visual reasoning. The source lists 428B total parameters, roughly 23B activated parameters and a 1M token context, and mentions sparse attention and a native vision encoder, with inputs supporting text, images and video and outputs being text. These total parameters match MiniMax M3, and the source notes that MiniMax has not released a technical report explaining the specific changes in the Flash variant. The report says that at the time of reporting the model was available on the MiniMax API, M Plan and the MiniMax Code desktop client, with Anthropic-compatible and OpenAI-compatible access (including Chat Completions and Responses). The report gives unified pricing: $0.15 per million input tokens, $0.60 per million output tokens, $0.015 for cache reads, with no long-context surcharge; the source says this is roughly half the per-token standard M3 rate. Reasoning is always on and cannot be disabled, with five effort levels: Low, Medium, High, Xhigh and Max, and MiniMax recommends Max as the default; sending thinking.type=disabled or effort=none returns an error, and reasoning content is placed in a reasoning_content field that may be absent. The source says weights are expected to be released on October 23, 2026, that the license is undisclosed and that self-hosting hardware is undisclosed; it also says Space Bunny is an early build of M3.1 Flash. All benchmark scores are reported by MiniMax and have not been independently reproduced.
What it means for developers
For developers, the most immediate consequence is not the parameter scale but the change in the integration contract: reasoning cannot be disabled, meaning every request carries a reasoning-token and latency floor, which changes the cost structure for workloads such as simple classification, extraction and short question answering. The second change is the parsing path: the source explicitly states that integrations relying on extracting reasoning from inline think tags in M3 need to switch to reading the separate reasoning_content field, and that this field may be absent, so parsing code must tolerate its absence or it will treat normal responses as errors. The third is effort levels: the five settings turn the quality/cost trade-off from a model-selection problem into a parameter-selection problem, letting the same model on the same workload produce markedly different scores and spend; the source cites an Automation-Bench chart showing each level below $0.05 per task. The fourth is the compatibility layer: OpenAI-compatible and Anthropic-compatible access are both provided, meaning migration paths for existing SDKs are relatively broad, but effort and reasoning fields are vendor extensions that are hard to carry over unchanged when switching between providers. It should be emphasized that the benchmarks, costs and speed above (the report mentions about 150 tokens/second on the product page and a claim in the announcement of 3x M3) are all MiniMax-reported figures without independent verification; the source itself also records an inconsistency between the chart and the table (62.8 versus 62.2 for the maximum effort level) and notes that Automation-Bench numbers cannot be compared directly across sources, for example the Kimi K3 model card reports 30.8 on the public subset of 600 tasks while the MiniMax table lists 46.7. Coding ability is the weakest part of this narrative: the source says Terminal Bench 4.0 is 38.1, behind Opus 5's 51.8 by about 13.7 points, and NL2Repo-Bench is also behind by about 14.1 points. On self-hosting, until the weights and license are actually published, deployers cannot judge commercial and redistribution rights; the source only notes that the MXFP8 build of M3 is about 440 GB, which can serve as a rough reference for hardware scale but is not the actual size of M3.1 Flash.
Integration and evaluation
Recommendation one: first verify the parsing changes for the reasoning field before discussing migration. The reported change triggering this decision is that reasoning is always on, disabling it returns an error, and the content moves into reasoning_content. The practical choice is to change the internal reasoning extraction layer from matching inline tags to reading a structured field and allowing it to be absent. The precondition is that parsing or post-processing logic relying on inline tags exists in the codebase; if parsing already reads a structured field, this work can be skipped. A measurable verification criterion is to count, on a fixed sample set, the response success rate and parse failure rate when reasoning_content is absent, confirming that absence does not cause content loss or exceptions. Recommendation two: treat effort levels as a tunable parameter and produce a cost-quality curve, rather than copying the officially recommended Max. The reported change is five effort levels with Max recommended as the default. The practical choice is to fix a level per workload type and write it into configuration, rather than using one global setting. The precondition is that the task has an automatically judgeable success criterion; for open-ended tasks without one, small-sample human labeling should come first. A measurable verification criterion is the pass-rate gap on the same sample set between Low and Max, along with the corresponding changes in token consumption and end-to-end latency; if the gap is below the noise level, choose the lower level. Recommendation three: decide on self-hosting after the weights are released, and do not do capacity planning around 440 GB in advance. The reported change is that neither the weights nor the license are disclosed. The practical choice is whether to continue using the API or to evaluate local deployment once the license is clear. The preconditions are whether the license permits the intended use and whether the hardware meets the actual weight size; until both are clear, a self-hosting evaluation lacks a basis. Measurable verification criteria are whether the license's terms on commercial use and redistribution cover your use case, and the measured throughput of the available quantized version on your target GPU.
Limits and open questions
There are three key uncertainties. First, all benchmark, cost and speed data are MiniMax-reported with no independent reproduction; the source itself records the inconsistency between 62.8 and 62.2 in the same announcement and the non-comparability of Automation-Bench across sources. Second, the weights are planned for release on October 23, 2026, but as of the report both the license and self-hosting hardware were undisclosed; weight availability does not equal a license granting commercial or redistribution rights, and the two must be judged separately. Third, MiniMax M3.1 Flash does not currently appear in the KimiSeek live model catalog; the API availability and pricing described by the source refer to MiniMax's official channels. If calling through KimiSeek is later considered, the actual entries on the site's pricing page and model list should be taken as authoritative, and prices may change. In addition, the source provides no technical report on Flash relative to M3, so the specific implementation of sparse attention and details of the vision encoder cannot be verified. The integration recommendations and verification methods in this article are analysis based on reported changes, not verified conclusions about the model's actual performance.