News · 2026-10-20

MiniMax launches M3.1 Flash as a cheaper agent model, with weights still to come

MiniMax launched M3.1 Flash on October 19 as a low-cost hosted model for software and business agents, charging $0.60 per million output tokens. The company says downloadable weights will follow, but its announcement does not confirm the October 23 date circulating on Reddit or establish a model-specific license. The release expands cheap agent options without yet establishing dependable autonomy or practical local deployment.

Key facts

• MiniMax's standard rates are $0.15 per million uncached input tokens and $0.60 per million output tokens.
• The October 19 release is hosted; MiniMax says, “Model weights will be released later.”
• The preview guide lists roughly 428 billion total parameters, with about 23 billion active for each token.
• The primary source is MiniMax's release announcement, supported by its preview guide.

The significant price is the cost of completed work, rather than an inexpensive paragraph. MiniMax pitches the model for reading a repository, making a change, running tests and continuing until a software task is done. Its business examples span sales, finance, operations and customer support. Those workflows consume many model calls, so a small price difference can accumulate across an agent's attempts, corrections and tool results.

An agent model is only one component of that process. Think of it as the planner at a workbench: the surrounding application supplies the tools, decides which drawers it may open and records what happens. A model does not gain access to a company's accounts merely because its developer advertises business-agent capabilities. The agent harness, permissions and independent checks still determine whether a plausible plan becomes a correct result.

MiniMax's preview guide describes a million-token context window, image and video input, text output and always-enabled reasoning with selectable effort. Those settings matter for migration: a developer should compare the same workload and effort level, rather than assume a faster label guarantees faster completion. The guide itself describes its case studies as examples, not guarantees of what every customer will obtain.

The model's size also needs careful interpretation. In a mixture of experts, only some internal specialists contribute to each token, but the wider collection still exists. The roughly 23-billion-active figure describes computation; it is not the size of a download. No verified M3.1 checkpoint file total was available in the dossier, so an exact disk size cannot be stated. The primary material also leaves the local graphics-memory requirement unstated. Neither number should be reconstructed from the parameter count.

The official M3 repository supplies useful family context, including sparse attention, but it does not establish M3.1's eventual license or complete implementation. MiniMax says the new release uses more aggressive sparsity. That can reduce attention work without eliminating the storage needed for the full set of expert weights. A promise to release weights therefore leaves several practical questions open: the actual files, supported inference code, terms of use and measured hardware configuration.

The reception is more mixed than the launch charts suggest. The Hacker News discussion includes strong complaints about real-world coding quality alongside reports of worthwhile results. Some comparisons involve an earlier preview checkpoint, so those experiences cannot simply be assigned to the final release. Different harnesses and reasoning settings also make casual comparisons difficult.

Release identity also matters. Ground Truth’s coverage of DeepSeek’s production refresh shows why a familiar model alias can conceal a changed checkpoint. A fair comparison records the actual version and evaluation settings, rather than relying on a family name alone.

The most consequential calibration warning concerns a reported 93% hallucination figure. VentureBeat attributed that reading to Artificial Analysis, but the dossier could not recover the exact model point from the evaluator's readable chart. What it did verify is the evaluation's definition: the denominator is responses that were not fully correct. The figure would describe how often those responses were wrong rather than partial or abstaining; it would not mean that 93% of all answers were false.

That distinction matters when factual uncertainty feeds an action. A cheap model that confidently fills an unknown field can create more cleanup than one that declines. Meanwhile, Epoch AI's audit of DeepSWE found benchmark problems, weakening one highlighted coding comparison without proving MiniMax's reported score false. The useful conclusion is specific: M3.1 Flash is a shipping, inexpensive agent candidate. Its suitability still depends on task-level evaluation, honest uncertainty handling and checks around the actions that matter.

Primary source, verified: read the paper →

Ground Truth. (2026, October 20). MiniMax launches M3.1 Flash as a cheaper agent model, with weights still to come. Ground Truth. https://groundtruth.day/news/minimax-m31-flash-cheaper-agent-api-unconfirmed-weight-date.html

@misc{groundtruth:minimax-m31-flash-cheaper-agent-api-unconfirmed-weight-date,
  title  = {MiniMax launches M3.1 Flash as a cheaper agent model, with weights still to come},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/minimax-m31-flash-cheaper-agent-api-unconfirmed-weight-date.html}
}

Topics: minimax · models · agents · inference-cost · open-weights