MiniMax hasn't been a household name in the reasoning model race, but the M1's track record shows they are serious about the work.
M1 Positioning
The M1 is MiniMax's reasoning model series, competing directly with OpenAI's o-series and DeepSeek's R1. Its core strengths lie in mathematical reasoning, code generation, and logical analysis.
On SWE-bench Verified, MiniMax M2.5 (an iterative version of M1) scored 80.2% — placing behind Claude Opus 4.5 and Gemini 3.1 Pro, but 0.2 percentage points ahead of GPT-5.2. This result ranks among the top tier globally.
MoE Architecture
The M1 series also uses the MoE (Mixture of Experts) architecture, following the same path as DeepSeek and Qwen. This approach maintains a high parameter count (large capacity) while controlling computational costs during actual inference.
Combined with Lightning Attention technology, MiniMax offers notable cost efficiency in long-context reasoning scenarios — consuming less compute and time for the same tasks.
Low-Key Strategy
Compared to other domestic large-model companies, MiniMax keeps its PR very restrained. It is not keen on publishing benchmark scorecards or marketing articles, and product updates rarely involve press conferences.
This approach has both pros and cons: the upside is that it avoids a "bragging" impression, while the downside is low market awareness. Many developers are unaware that MiniMax's models can already compete head-to-head with top players.
Commercialization Path
MiniMax's commercialization follows a To C + API dual track:
- To C: Hailuo AI targets general users
- API: Provides developer access interfaces
This strategy differs from both Anthropic (starting with pure API then adding To C) and OpenAI (ChatGPT first then API). MiniMax has walked on two legs from the start, likely because as a relatively small player, it needs multiple revenue streams to support R&D.
Sources: SWE-bench leaderboard, CocoLoop, MiniMax official API documentation