MiniMax, a $20 Billion AI Unicorn, Launches Its First Inference Model M1, Surpassing Domestic and International Models Like DeepSeek
Industry Background and Release Overview
The global artificial intelligence sector is currently experiencing an unprecedented wave of development, particularly in the area of large language model inference technology. Major tech giants and startups are ramping up their investments. Following the early releases of inference models by companies like OpenAI and DeepSeek, industry leaders such as Alibaba, Baidu, Tencent, ByteDance, and Google have also launched their own solutions. In this round of technological competition, latecomers often adopt a strategy of 'accumulating strength before striking,' achieving breakthroughs through technological innovation.
Against this backdrop, MiniMax—a unicorn valued at over 20 billion yuan—recently officially released its first inference model M1. According to official disclosures, M1 is the world's first large-scale mixed attention inference model that uses open-source weights; it marks another significant breakthrough for Chinese AI companies in core technology fields. This release is the first major achievement during MiniMax's five-day technical launch week; subsequent updates will include intelligent applications and enhancements to Hai Luo AI in video and music domains.
Model Performance and Technical Breakthroughs
MiniMax M1 is a new inference model developed based on the MiniMax-text-01 architecture with several innovative features regarding its technical structure and performance. The total parameter scale reaches 456 billion with 45.9 billion active parameters per token; it employs cutting-edge technologies including Mixed Experts (MoE) architecture and Linear Attention Mechanism (Lightning Attention). For evaluation purposes within the industry’s mainstream assessment sets comprising 17 benchmarks overall testing showed that it has reached or surpassed leading domestic and international models across multiple key metrics.
Specifically speaking about professional areas like mathematical reasoning or code generation where M1 excels notably: test data indicates that its performance exceeds even Anthropic's strongest Claude-4-Opus model while outperforming ByteDance's latest Seed-Thinking-v1.5 as well as DeepSeek-R1 which boasts 671 billion parameters too—but it's important to note there remains some gap when compared against versions like DeepSeek-R1-0528 alongside OpenAI’s O3 model or Google's latest Gemini 2.5 Pro.
Analysis of Professional Scene Advantages
In terms of application performance within specialized productivity scenarios,Mini Max demonstrates unique competitive advantages especially evident within complex use cases involving software engineering,long context comprehension,and tool utilization.For instance,in SWE-bench tests assessing software engineering capabilities,M achieved over55%excellent results although still lagging behind top overseas models yet significantly ahead than domestic counterparts such asDeep Seek R - one &similar offerings fromAlibabaandByte Dance . Long-context understanding capability represents another highlight feature forM . Across three authoritative benchmark tests ,it comprehensively outperformed all open-source models ,even surpassing closed-source commercial ones likeO3andClaude -4despite only trailing slightly behindGemini2 .5Pro ranking second globally.This accomplishment showcasesM ’s exceptional ability when processing lengthy text information effectively. Within agent tool usage contexts duringTAU-bench assessments ,performance remained commendable.M ‘s scores exceeded60 %in aviation-related evaluations leading current open/closed source alternatives ;retail domain performances ranked just belowO3&Claude -4models but above those offered byDeep SeekorAliBabaorByte Dance making these outstanding achievements placeM among most practical value-laden large-model options available today .
Architectural Innovations & Technical Details
m ’s excellence stems from various innovations introduced byMini Maxat both architectural level algorithmic aspects alike.Currently facing challenges related primarily towards howto overcome exponential growth computational requirements tied specifically into Transformer architectures’ attention mechanisms dependent upon sequence lengths.To address these issues seen throughout industry firms includingDeep SeekMoonlitDarksidehave proposed native sparse attentions(NSA)block-attention hybrid structures(MoBA). nHowever,minimax chose alternative routes innovatively adopting linear attentional methods(LightningAttention )based offMixedExpertarchitecture.The core idea revolves around decomposingattention calculationsinto smaller chunks employinglinear complexity computation approaches thereby facilitating efficient handlingof long sequences.BasedonMinimax’stechnical papers,this design theoretically allowsfor effective extensionof reasoning length reaching hundreds thousands tokens whilst drastically reducingcomputational costs involved.In practice,test results reveal substantial efficiency gains versus competitors:when generating64ktokens,mconsumed less than50%FLOPScomparedagainstdeepseek-r-one ;100k tokenlength consumption dropped down approximately25%.Such notable improvements provide crucial support enablinglonger contextual processing thus rendering ideal fitfor tasks requiring deep thinking intricate inputs simultaneously addressingreal-world complexities efficiently . n### Reinforcement Learning Training Optimizationm success owes much thanks duelargescale reinforcement learning algorithms pioneeredbyMinimax.Two key innovations highlightedwithinits technical documentationinclude introduction novelCISPOreinforcementlearningalgorithmwhich greatly enhances trainingefficiency.A comparative experiment conductedusingmathematicaltestingbenchmarkAIME revealedthatCISPOachievedtwice accelerationoverbyte dance recent DAPOalgorithm needingonlyhalfas manytrainingsteps attain comparableperformancesignificantly outperformingpreviousGRPOapproaches utilizedbyDeep SeekMoreover,specializedsolutionswere devised tacklingaccuracy mismatches encounteredduringhybridarchitecturalexpansionprocesses stemmingfromdifferences betweentrainingkernelsinferredkernelsthat impede rewardgrowthduringreinforcementlearningperiod.Minimizingaggressiveextensionstrategiesavoidinggradientexplosion riskswas ensuredvia phased progressivecontext-length expansion plans dividingtheentiretrainingsequenceinto four stages starting32k gradually extendingupwards untilreachingone milliontoken lengths ensuringstablemodel-training conditionsoverallbenefiting fromtheseinnovations resultedhighlyefficient reinforcement training processes reportedwith512H800GPUsoperatingthree weeks costingapproximately$537400equivalent3800000RMBsignificantly lowerthaninitialexpectationsdemonstratingeconomic benefits derivedtechnological advancementsmade possible throughthis approachmoving forward towardcommercialization strategies market positioningwill be essential factors determiningfuture trajectoriessuccessfully navigatingcompetitive landscapes while capitalizinguponopportunities arisingwithinemergingfieldsrelatedartificialintelligenceapplications.Mini maxcurrently pursuesactive marketingstrategiesregardingcommercialapplicabilityoftheirproducts presently offeringfree upgrades viaAPI pricing modes drawing inspirationfrombyte dancesinterval-basedpricing frameworks tailoredaccordingdifferentinputlengthranges deliveringcost-effectiveoptions relativecompetitionprovidingsolutions spanningacrossvariousdomainsincludingaviationretailamongothers showcasingremarkablepotential unlockingvalue-driven prospects amidst evolvingecosystems shapingtomorrow’stechnology landscape.