MiniMax Unleashes M2
Shanghai-based MiniMax released MiniMax M2; a 10B-parameter "Giant Killer" billed at 8% the cost of Claude Sonnet.
In a move set to disrupt the AI industry, Shanghai-based MiniMax has announced the release of its MiniMax-M2 model. The launch accelerates a market-wide shift from high-cost, closed-source systems toward powerful, open-source alternatives. MiniMax claims M2 can generate tokens "nearly as good as Claude Sonnet 4.5" at twice the speed and for only 8% of the cost. This aggressive strategy puts direct pressure on the pricing models of major international tech companies. The model was specifically built to excel at AI Agent tasks and code generation, positioning it as a foundational base for building the next generation of complex, multi-step AI applications.
Performance and Price Engineering
MiniMax's claims are supported by external benchmarks. An evaluation from Artificial Analysis identified MiniMax-M2 as the top-performing open-weights model currently available, placing it in the "global top five" among all models, including closed-source ones. M2 achieves an average intelligence index of 61%, a score that achieves crucial competitive parity with the very closed-source models it aims to disrupt. This score surpasses Google's Gemini 2.5 Pro (60%) and is nearly identical in performance to Anthropic's Claude Sonnet 4.5 (63%). This means developers can now access performance on par with top-tier models at a 12.5-fold reduction in cost.
The "How-It-Works" Deep Dive
M2's revolutionary price-to-performance ratio is the direct result of three key innovations: its architecture, its interaction protocol, and its training method.
1. Efficient MoE Architecture
M2's efficiency stems from its Mixture-of-Experts (MoE) architecture. It contains 230 billion total parameters, but only 10 billion active parameters are used to generate a token. This is significantly more efficient than rivals like DeepSeek, which activates roughly three times as many parameters per token. This lean design is the source of M2's speed and cost-effectiveness.
2. Interleaved Reasoning
M2 is explicitly defined as an "interleaved thinking model". This refers to a critical data protocol where the model's internal "thought process" is externalized and wrapped in XML-style tags: <think>...</think>.
The model's documentation warns developers that they must retain this thinking content in the conversation history. Removing the <think>...</think> blocks will degrade the model's performance. This process allows the model to build upon its own reasoning, enabling it to self-correct and manage the complex "plan → act → verify" workflows required by advanced AI agents.
3. CISPO: The "Secret Sauce" for Training
M2's disruptive economics are enabled by an internal innovation in its training process: CISPO , which stands for "Clipped IS-weight Policy Optimization". This is a novel and "faster reinforcement learning (RL) algorithm" developed by MiniMax.
This algorithm diverges from traditional methods; it "clips the importance sampling weights" to make training more stable and rapid. MiniMax claims CISPO's convergence is "twice as fast as other RL algorithms". This superior manufacturing process dramatically cuts training costs. For instance, the M1 model's entire RL phase cost only $534,700. This internal efficiency is what allows MiniMax to offer M2 at such a low price.
A Three-Pronged Strategy for Adoption
MiniMax is rolling out M2 with a comprehensive strategy designed for maximum, frictionless adoption :
For Researchers: The complete model weights are open-sourced on Hugging Face, allowing for local deployment and academic research.
For Developers: A commercial API is on the MiniMax Open Platform, offered "free for a limited time". Its standard pricing is set at a disruptive $0.30 (input) and $1.20 (output) per million tokens.
For End-Users: A free, M2-powered MiniMax Agent product has been released, serving as a public demonstration and user-acquisition tool.
By open-sourcing a model with elite performance at a fraction of the cost, MiniMax has fundamentally altered the market's "pricing logic" and fired the starting gun on the large-scale commoditization of top-tier intelligence.
