Correction — October 3, 2026: We removed architecture and AST-execution claims inferred from a model selector, along with unsupported performance claims. The retained graphic illustrates interface labels; it is not an independent benchmark or evidence of the model’s internal design. Access information was checked against MiniMax’s current plan page.
MiniMax’s M Plan page lists M3.1 Flash Preview as a text model across its Go, Explore and Build subscriptions. It offers MiniMax Code and an API key for compatible tools. A product’s availability is useful evidence about access, but it does not establish how its weights, routing or reasoning implementation work.
The selector is a control surface
The graphic shows low, medium, high, xhigh and max effort choices. Labels communicate the product’s controls. They do not disclose a fixed number of reasoning tokens, guarantee a latency, or prove that the system directly manipulates an abstract syntax tree. The behaviour of the model, the surrounding agent and its tools must be examined separately.
For example, an agent editing a source file can use text replacement, a patch tool or a language-aware editor. Seeing a successful code change does not identify which method the model itself implements. The application trace and tool documentation are better evidence than the presence of a model name in a dropdown.
Check the subscription key and credit rules
The plan page distinguishes a Subscription Key from a pay-as-you-go API key. MiniMax says the former uses the plan allowance and then available credit packs, while the latter can charge an account balance. Its current table lists limited free text-model access; paid tiers list M3.1 Flash Preview. Video-model access differs by tier.
An integration should check the key type, model identifier and the account’s current usage page before an unattended job starts. Promotional credit resets are not a permanent capacity commitment. This revision removes the earlier implication that a short campaign establishes an ongoing entitlement.
Hosted availability also differs from permission to download and deploy model weights. Our Ling 3.1 Flash access guide explains that distinction for another hosted Flash model. A model’s name or active-parameter count cannot, on its own, establish a local deployment path.
Choose effort from a repeatable task
A useful evaluation uses the same repository revision, instructions, tool permissions and acceptance check at two or three effort settings. Record elapsed time, billed usage, successful completion and required human edits. Repeat representative tasks so that a single favourable run does not become the recommendation.
Start with a task whose output you can verify: a failing test to repair, a small interface change or a bounded document extraction. Judge the final artifact, not how long the model appears to think. If higher effort produces more work outside the requested scope, that is a reason to adjust instructions or effort for that task rather than declare a universal ranking.
This article establishes an access route and an evaluation approach. It does not report independent M3.1 latency, parameter counts, benchmark superiority or a verified internal reasoning architecture. Those claims require a model specification or reproducible results beyond a product selector.
