MiniMax M3 Debuts With 9.4X CUDA Acceleration and Autonomous Model Training
Chinese AI startup MiniMax launched its next-generation flagship model M3 on June 1, 2026, introducing a breakthrough sparse attention architecture that delivers up to 15X faster decoding speeds while autonomously training four models from scratch in a 12-hour window—without human intervention.
The Beijing-based company, currently pursuing an A-share IPO after filing with the Shanghai Securities Regulatory Bureau, saw its Hong Kong-listed shares drop 14.64% to HK$717.00 (US$92.19) following the announcement, bringing market capitalization to HK$224.88 billion (US$28.90 billion). The stock decline suggests investor caution despite the technical advances, possibly reflecting concerns over monetization timelines in China's increasingly competitive large language model (LLM) sector.
MiniMax M3's performance edges past GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro—a benchmark measuring real-world software engineering capabilities—though it trails Claude Opus 4.7. The model's distinguishing feature lies in its MiniMax Sparse Attention (MSA) architecture, enabling a 1-million-token context window with computational efficiency gains of 9X during prefilling and over 15X during decoding compared to its predecessor M2's full-attention mechanism.
Sparse Attention Redefines Efficiency-Performance Tradeoff
The MSA architecture addresses the quadratic complexity growth inherent in traditional full-attention mechanisms by implementing a precision-oriented KV block partitioning strategy. Unlike existing sparse attention solutions such as DSA and MoBA, MSA achieves higher effective context coverage through more accurate block division, according to MiniMax's technical disclosures.
At the operator level, MiniMax engineers adopted a "KV outer gather Q" strategy that clusters queries by KV block as the outer loop—ensuring each block is read only once with continuous memory access. Under M3's current head configuration, this design delivers a compute-to-memory-access ratio significantly superior to mainstream methods, running over 4X faster than open-source implementations like Flash-Sparse-Attention and FlashMoBA.
At 1-million-token context length, M3's per-token computation drops to just 1/20th of the previous generation. Controlled experiments demonstrate MSA matches full-attention performance across most capability dimensions while dramatically cutting inference costs—a critical factor as Chinese LLM providers face margin pressure from price competition.
Native Multimodal Training Scales to 100 Trillion Tokens
M3 represents MiniMax's first model trained with multimodal data from "Step 0," mixing text, images, and video throughout pretraining rather than bolting vision capabilities onto a text-only foundation. The company's internal experiments revealed that interleaved data—sequences where text and visual modalities alternate naturally—proves more crucial to performance gains than previously recognized in industry practice.
After reconstructing its entire data pipeline to optimize for interleaved inputs, MiniMax scaled training corpus to 100 trillion tokens. On OmniDocBench, a multimodal evaluation suite, M3 outscored Gemini 3.1 Pro. The model also topped Claw-Eval, an end-to-end framework assessing autonomous agent capabilities, including computer desktop operation.
However, hands-on testing by Chinese tech media revealed mixed practical results. While M3 generated aesthetically acceptable racing game interfaces, playability remained limited. In dynamic SVG generation tasks, the model produced rough approximations of bicycles and pelicans but struggled with accurate visual representation. Location identification from photos yielded detailed scene analysis—noting high-rise buildings and dust nets—but failed to pinpoint geographic specifics, a task where competitors like DeepSeek and Qwen3.7 Max Preview achieve 10-kilometer accuracy.
Autonomous Post-Training Demonstrates Emerging Self-Improvement
M3's most striking capability surfaces in PostTrainBench, a benchmark evaluating post-training competence. Given four pretrained base models and a 12-hour deadline, M3 autonomously executed the complete loop of data synthesis, training, evaluation, and iteration—no human oversight—bringing the quartet to basic proficiency across five tasks including mathematical reasoning, tool invocation, and code generation. M3 scored 0.37, trailing GPT-5.5 (0.39) and Opus 4.7 (0.42) but substantially ahead of other models.
In a separate 24-hour session involving 147 benchmark submissions and 1,959 tool calls, M3 optimized a CUDA kernel from 7.6% to 71.3% utilization of Hopper FP8 hardware peak performance—a 9.4X speedup over baseline. This self-directed optimization underscores the model's potential in automating traditionally expert-intensive tasks, though scaling such capabilities to production environments remains unproven.
MiniMax simultaneously updated its AI coding tool MiniMax Code, co-designed and co-trained with M3 to leverage the model's long-context, agentic, and multimodal strengths. The tool features computer-use functionality, executing cross-application, cross-file, and cross-system operations on desktop environments. MiniMax plans to open-source both M3 model weights and MiniMax Code within the next 10 days, following release of a technical report.
Pricing Strategy Targets Enterprise Adoption
M3 API access splits into two tiers based on a 512K-token context threshold, offering priority and standard invocation modes. A seven-day 50% discount applies to sub-512K contexts, with standard pricing at RMB 2.1 (US$0.29) per million input tokens and RMB 8.4 (US$1.17) per million output tokens. Cache reads cost RMB 0.42 (US$0.06) per million tokens—positioning M3 competitively against ByteDance's Doubao and Alibaba's Qwen offerings in China's price-sensitive market.
The model's interactive user simulator framework—designed to mirror real developer collaboration patterns including requirement clarification, iterative refinement, and continuous task switching—aims to close the gap between benchmark performance and production experience. Testing confirmed M3 grasps collaborative workflows and autonomously iterates, though task completion quality remains inconsistent. In one trial, M3 drafted a comprehensive plan for a Google Docs-style collaborative system MVP, accounting for technical stack, unit testing, and risk mitigation, but execution remained incomplete at publication time despite active self-correction and continuous optimization.
As Chinese LLM developers intensify exploration of architectural innovation beyond incremental parameter scaling, MiniMax's dual bet on sparse attention and native multimodal training signals an industry pivot toward efficiency-first design. Whether these technical advances translate to defensible market position amid fierce competition from Alibaba, ByteDance, and Baidu—each commanding deeper enterprise distribution channels—will determine MiniMax's IPO reception when it formally lists on mainland exchanges.
Related Coverage：
Zhipu AI's 10x Rally Exposes Hong Kong's AI Narrative Premium Over MiniMax