MiniMax M3 AI model targets coding agents, beats GPT-5.5 on SWE-Bench
Synopsis
MiniMax's new M3 model claims to beat OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro on the SWE-Bench Pro coding benchmark, while cutting inference costs to one-twentieth of previous levels — and it arrives just as the company gears up for a Shanghai Star Market IPO.
Key Takeaways
MiniMax launched its M3 flagship AI model on Monday, 1 June 2026 , targeting coding agents and automated software workflows.
M3 's redesigned architecture cuts computational requirements to as little as one-twentieth of previous levels, reducing inference costs significantly.
The model supports a context window of up to 1 million tokens — five times the capacity of its predecessor, M2.7 .
According to MiniMax , M3 outperformed OpenAI GPT-5.5 and Google Gemini 3.1 Pro on the SWE-Bench Pro coding benchmark.
In a company-cited test, M3 successfully optimised software running on Nvidia 's Hopper GPU architecture.
The launch is MiniMax 's first major product release since the firm began preparations for an IPO on Shanghai 's Star Market , supplementing its existing Hong Kong listing.
Chinese AI start-up MiniMax has launched M3, its latest flagship model engineered for long-context coding tasks and automated software workflows, marking the company's most significant product release since it began preparations for a dual listing on Shanghai's Star Market. The Shanghai-based firm announced the model on Monday, 1 June 2026, positioning it as the foundation of its push into coding agents and enterprise automation.
Architecture overhaul slashes inference costs
According to the company, M3's redesigned architecture reduces computational requirements to as little as one-twentieth of previous levels, delivering sharply lower inference costs alongside faster response speeds. The efficiency gains are significant for enterprise customers who run high-volume, latency-sensitive workloads. MiniMax did not disclose the model's parameter size or the computing infrastructure used during training.
One million token context window, five times its predecessor
M3 can process up to 1 million tokens of data in a single context window — five times the capacity of its predecessor, M2.7 — enabling it to ingest and reason over entire large-scale codebases in one pass. In a benchmark test cited by the company, M3 successfully identified optimisation strategies for software running on Nvidia's Hopper chips, a widely used GPU architecture in AI data centres.
Outperforms OpenAI and Google on SWE-Bench Pro
According to MiniMax's official WeChat post, MiniMax-M3 surpassed rival models including OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro on SWE-Bench Pro, a leading industry benchmark for software engineering capability and automated task completion. The result positions M3 as a competitive entrant in the rapidly crowding field of AI coding assistants, where players such as Anthropic and DeepSeek are also vying for developer mindshare.
Why it matters: IPO timing and market positioning
The M3 launch is the first major product milestone since MiniMax formally initiated preparations for an initial public offering on Shanghai's tech-heavy Star Market, which would complement its existing listing in Hong Kong. Demonstrating frontier-level coding performance ahead of a public market debut signals the company's intent to compete directly with both Western AI labs and domestic rivals. Investors and analysts will be watching whether the efficiency and benchmark claims translate into enterprise adoption and recurring revenue.
What's next
With coding agents emerging as a primary commercial battleground for AI companies globally, MiniMax's next moves — particularly around developer tooling, API pricing, and enterprise partnerships — will determine how quickly M3 converts benchmark wins into market share. The company's dual-listing trajectory on the Star Market and Hong Kong exchanges adds a financial urgency to sustaining this product momentum.
Point of View
It is the frame. What mainstream coverage underweights is the architecture story: a 20x reduction in compute per inference, if independently verified, matters far more commercially than leaderboard position, because it directly attacks the unit-economics problem that is slowing enterprise AI adoption globally. The 1 million token context window also puts MiniMax in direct competition with long-context plays from Anthropic and Google, not just domestic rivals like DeepSeek. Notably absent from the announcement is any detail on training compute or model size — omissions that will matter to investors and enterprise buyers evaluating reproducibility and fine-tuning costs.
NationPress
19 Jul 2026
Frequently Asked Questions
What is MiniMax M3 and what makes it different?
MiniMax M3 is a new flagship AI model from Shanghai -based start-up MiniMax , designed specifically for long-context coding tasks and automated software workflows. Its redesigned architecture cuts inference compute to as little as one-twentieth of prior levels and supports a 1 million token context window, five times larger than its predecessor M2.7 .
How does MiniMax M3 compare to OpenAI GPT-5.5 and Google Gemini 3.1 Pro?
According to MiniMax 's own WeChat announcement, M3 outperformed both OpenAI GPT-5.5 and Google Gemini 3.1 Pro on SWE-Bench Pro , a major industry benchmark for software engineering and automated task performance. The claims have not been independently verified.
Why is MiniMax launching M3 now?
The M3 launch is MiniMax 's first major product release since the company formally began preparations for an IPO on Shanghai 's Star Market , which would complement its existing Hong Kong listing. A strong product showing ahead of a public market debut helps establish competitive credibility with institutional investors.
What is SWE-Bench Pro and why does it matter?
SWE-Bench Pro is a widely used industry benchmark that evaluates an AI model's ability to handle real-world software engineering tasks and automated workflows. Strong performance on this benchmark is considered a reliable signal of a model's practical utility for coding agents and developer tools.
Which companies does MiniMax compete with in AI coding?
MiniMax competes in the AI coding space with global players including OpenAI , Google , and Anthropic , as well as Chinese rivals such as DeepSeek . The coding-agent segment is one of the fastest-growing and most commercially contested areas in enterprise AI.
