Official Launch and the 428B Sparse MoE Architecture
On October 19, 2026, MiniMax officially unveiled its latest foundation model, M3.1 Flash. The lab confirmed that the release is the official identity of the stealth evaluation checkpoint previously operating under the moniker 'Space Bunny,' which had quietly processed over 200 trillion tokens prior to its public branding.
Architecturally, M3.1 Flash is engineered as a multimodal Mixture-of-Experts (MoE) model comprising 428 billion total parameters, but it routes each token through approximately 23 billion active parameters. Featuring sparse attention mechanisms alongside a native visual encoder, the model supports an input context window of 1,000,000 tokens and an output capacity of up to 524,288 tokens.
Always-On Reasoning Mechanics and API Cost Economics
A distinctive operational constraint in M3.1 Flash is its always-on reasoning pipeline. Reasoning cannot be disabled; calls configured with thinking.type=disabled return runtime errors. The system exposes five distinct effort levels—Low, Medium, High, Xhigh, and Max—and pipes intermediate cognitive chains through an explicit reasoning_content response field.
The model is currently accessible via the MiniMax API under MiniMax-M3.1-Flash, within the MiniMax Code desktop application, and through the M Plan subscription tier. MiniMax set API commercial rates at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. Furthermore, the company announced that full model weights are scheduled for public release on October 23, 2026.
Reported Performance Benchmarks and Verification Realities
According to self-reported vendor evaluations, MiniMax documented an Automation-Bench v1.0.6 score of 62.2 (rising to 62.8 at max effort), a BankerToolBench score of 76.2 matching Claude Opus 5, and a DeepSWE v1.1 mark of 73.7%. On the Agents' Last Exam (ALE) benchmark, it registered 29.6% across 152 tasks for an estimated cost of approximately $111, alongside an OSWorld 2.0 rating of 61.7, CyberGym score of 82.2, and Terminal Bench 4.0 score of 38.1.
These figures, however, remain self-reported vendor claims that have not yet undergone reproduction by independent benchmark registries. Sweeping claims that M3.1 surpasses all regional rivals are also incomplete; MiniMax's own comparative materials show that DeepSeek-V4.1-Flash-1017 retains a distinct lead on the DeepSWE benchmark (77.0% versus 73.7%).
Community and Practitioner Reaction: Efficiency Versus Contamination Concerns
Among software engineers and AI practitioners, early reaction has heavily favored the model's speed and aggressive cost efficiency. The sparse routing allows throughput figures ranging between 170 and 200+ tokens per second, with per-task execution costs dropping below $0.05, making it an appealing engine for autonomous agent loops.
Conversely, technical observers have raised skepticism regarding potential benchmark contamination, noting that the reported scores appear suspiciously pristine. Practitioners observed that during the unbranded 'Space Bunny' trial period, coding harness evaluations frequently favored alternative systems like GLM-5.3-Flash and DeepSeek-V4.1-Flash for production software tasks. Unsubstantiated claims that weights were immediately available for local workstation hardware were also quickly debunked, as local deployment remains blocked until the official release on October 23.
Implications for Thai Enterprises and Practical Strategy
For enterprise IT leaders and startups in Thailand, M3.1 Flash significantly lowers the total cost of ownership for long-context workloads. A 1M context capacity combined with $0.015 per million token caching pricing makes enterprise document ingestion—such as multi-year corporate regulatory filings, complex loan contracts, and policy manuals—commercially viable at scale.
However, the enforced always-on reasoning loop creates runtime overhead that may prove unsuitable for immediate, low-latency applications like customer-facing chatbots. Thai engineering teams should evaluate M3.1 Flash for asynchronous agentic workflows, complex technical parsing, and internal data analysis, while preparing governance and infrastructure frameworks to audit the open weights once released later this week.
By combining near-frontier agentic performance with sub-$0.05 per-task execution costs and an open-weights timeline, M3.1 Flash introduces enterprise-grade reasoning economics that challenge incumbent proprietary labs.