MiniMax's M3 Flagship Model and Its Indicator Impress but Spark Divisive Backlash

The race for large language models continues to intensify, with AI startup MiniMax recently releasing its new flagship model, M3. According to benchmark results from the technical report, the model delivers impressive performance: scoring 59% on tests that simulate real-world software engineering tasks, surpassing GPT-5.5 and approaching Opus4.7. It also offers million-token context processing and native multimodal capabilities. However, despite these strong technical achievements, the launch has triggered significant backlash from the developer community, with particularly harsh criticism surfacing in Chinese online forums.

A primary source of skepticism revolves around the evaluation methodology. Technical details reveal that M3 used Claude Code—a competing model—as the evaluation framework for its coding capability tests. While it is common practice to use existing toolchains for agent evaluations, critics say that MiniMax essentially tested its own model against a competitor's framework and directly claimed high scores, comparing itself to the same competitor in public promotions. Many developers have labeled this approach disingenuous, as it makes it difficult to distinguish the model's native abilities from the enhancements provided by the framework.

Another point of contention is the sincerity of the open-source commitment. Unlike other vendors that release open-source models with full weights, MiniMax did not disclose M3's size or provide its weights. The company only stated it would open-source the model within ten days, offering only API access at launch. Since reproducibility and verifiability are core values of the open-source community, promoting open-source while withholding weights—though understandable from a commercial standpoint—has alienated a developer community that expects transparency and practicality.

What has frustrated heavy users most is the sudden change to the billing rules. Previously, MiniMax was known for its generous usage limits, restricting by request count without a monthly token cap. With the M3 release, the company introduced a new token-based plan that charges by total volume. While the official line is that the Plus plan offers strong token value, heavy users—especially those working with million-token contexts—find that each call consumes a large amount, quickly depleting package quotas. This has led to widespread complaints from long-time users.

Setting aside the operational controversies, M3 does have notable innovations at the architectural level. It introduces a self-developed sparse attention mechanism called MSA (MiniMax Sparse Attention), which mitigates the quadratic computational explosion of traditional Transformers by applying high-precision block partitioning and sparsification to key-value pairs. At the low-level operator level, the model adopts a novel aggregation computation method that improves memory access continuity, achieving speeds four times faster than the open-source Flash-Sparse-Attention. As a result, M3 boosts forward propagation and decoding speeds by 9x and 15x respectively under million-token contexts, reducing single-token computation to half that of the previous generation.