The dominant narrative in AI development has been one of scale: bigger models, more parameters, more compute, more capability. MiniMax, the Shanghai-based AI company, is betting that this narrative is incomplete — that the future of AI deployment lies not in ever-larger cloud-based models but in efficient, capable models that can run on the devices people already carry in their pockets.
The company's ABAB model, released on April 10, 2026, is the most compelling argument yet for this thesis. ABAB achieves performance comparable to GPT-4 on standard benchmarks while requiring only 7 billion parameters — roughly one-thirtieth the size of the models it competes with. This efficiency is not achieved through capability sacrifice but through architectural innovations that allow the model to do more with less.
The Architecture Behind the Efficiency
ABAB's efficiency stems from several architectural innovations that MiniMax's research team has developed over the past two years. The most significant is a novel attention mechanism that reduces the computational complexity of processing long sequences from quadratic to near-linear, allowing the model to handle context windows of up to 200,000 tokens despite its small size.
The model also employs aggressive knowledge distillation techniques, training the small model to replicate the outputs of much larger teacher models across a carefully curated set of tasks. This allows ABAB to benefit from the knowledge encoded in frontier models without requiring the compute to train at that scale.

Performance on Consumer Hardware
The practical implications of ABAB's efficiency are significant. The model can run at full speed on a smartphone with 8GB of RAM, achieving response times of under 500 milliseconds for typical queries. On a laptop with a dedicated GPU, it runs even faster. This means that applications built on ABAB can provide AI assistance without requiring a network connection — a significant advantage for privacy-sensitive applications and for users in regions with limited connectivity.
Data Visualization
ABAB vs. Cloud Models: Performance per Compute Unit
- ABAB (7B)
- GPT-4o (175B)
- Claude 3 Haiku
"The most powerful AI is not the one running in a data center — it's the one that's always with you, always available, and never sends your data to a server you don't control."
— Yan Junjie, CEO, MiniMax
Applications and Market Opportunity
MiniMax is targeting several specific market segments with ABAB. In consumer applications, the model enables AI assistants that work offline, AI-powered photo editing that processes images on-device, and real-time translation that doesn't require a data connection. In enterprise applications, ABAB enables deployment on edge devices in manufacturing, healthcare, and logistics environments where cloud connectivity is unreliable or prohibited for security reasons.
The healthcare opportunity is particularly significant. Medical devices that incorporate AI assistance — from diagnostic imaging tools to patient monitoring systems — face strict regulatory requirements around data privacy that make cloud-based AI difficult to deploy. ABAB's ability to run on-device means that AI assistance can be integrated into medical workflows without the data leaving the hospital's network.
MiniMax has already signed partnership agreements with several major smartphone manufacturers, including partnerships with two of the top five Android OEMs by global market share. These partnerships will see ABAB integrated into the default AI assistant features of hundreds of millions of devices over the next 18 months, giving the model a distribution advantage that is difficult for competitors to match.