MiniMax M3: The Open-Weight Model That Challenged GPT-5.5 and Shook the Market

MiniMax deliberately packaged M3 as a unified foundation model rather than a collection of specialist components. The architecture is natively multimodal, meaning it accepts and processes text, images, and video directly, without auxiliary pipelines bolted on after training. This integration also supports direct computer desktop operation—a capability that distinguishes it from models that treat multimodality as a post-hoc interface layer .

Parameter counts remain undisclosed, a strategic omission that has become common among Chinese frontier labs competing in the open-weight space .

Benchmarking Against GPT-5.5 and Gemini 3.1 Pro

MiniMax staked its credibility on public leaderboard results, claiming that M3 outperforms both GPT-5.5 and Gemini 3.1 Pro on core engineering and agentic benchmarks:

• SWE-Bench Pro (software engineering): 59.0%, above GPT-5.5 and Gemini 3.1 Pro, approaching Claude Opus 4.7 .
• BrowseComp (autonomous web browsing): 83.5, surpassing both GPT-5.5 and Gemini 3.1 Pro, and edging past Claude Opus 4.7 .
• SVG-Bench (vector graphics generation): scored above Claude Opus 4.7 .
• OmniDocBench (multimodal document understanding): MiniMax also reported strong performance here .
• Terminal-Bench 2.1 (terminal-agent tasks): 66.0% .
• Claw-Eval (end-to-end agentic evaluation): top score among tested models .

An important caveat accompanies these numbers: several results were generated using MiniMax's own infrastructure with agent-specific scaffolding, and independent verification had not been published at launch time . The company has committed to releasing open weights and a full technical report within approximately ten days, which would allow third parties to replicate these benchmarks or challenge them .

Pricing and Token Plans

MiniMax introduced a three-tier subscription structure for M3 access, called the Token Plan :

API pricing is configured separately, with input costs starting around $0.30 per million tokens under a temporary 50% launch promotion (standard rate $0.60 per million) and output costs approximately $1.20 per million tokens (standard $2.40 per million) . Billing is split by context length: calls under 512K tokens fall into a standard tier, while larger prompts that fill the 1M-token window attract a higher long-context rate .

With prompt caching optimization, the blended cost can drop as low as $0.06 per million tokens . These pricing levels position M3 as aggressively cheap relative to comparable closed-source frontier models, continuing the price-competition dynamic that has defined the open-weight ecosystem .

Open Weights: Immediate Access, Deferred Release

M3 is publicly positioned as an open-weight model, though MiniMax has not chosen to release the weights at the moment of launch. Instead, the company has pledged to publish the full model weights and a detailed technical report on Hugging Face and GitHub within roughly ten days of the June 1 announcement .

The phrasing—"open-weight" rather than "open-source"—has drawn criticism from some corners of the developer community, who note that withholding training data and recipes stops short of a full open-source release . The model is available immediately through MiniMax Code, the Token Plans, and the API, meaning developers can integrate it into workflows now even while waiting for the ability to host it themselves .

The Stock Market Reversal

MiniMax's shares (ticker 00100.HK) opened 5.24% higher on announcement day, rising to HKD 884 and briefly touching an intraday high of HKD 907.5 . The early optimism proved fragile. By the close of trading, the stock had not merely surrendered its gains but had fallen substantially further, ending the day down 15.7% (-132 HKD) .

Heavy short-selling accompanied the reversal, with approximately $185.33 million in short interest representing 5.93% of turnover . Analysts attributed the sell-off to profit-taking by investors who had already enjoyed an extraordinary run: since MiniMax's Hong Kong IPO in January 2026 (which raised roughly $619 million), the stock had surged over 400% by late May .

The M3 launch occurred alongside MiniMax's May 29 announcement that it had signed a sponsorship agreement with CITIC Securities to begin its A-share IPO process on the Shanghai STAR Market . The dual catalysts—new product plus a secondary listing filing—generated the morning spike but may also have created the conditions for a "sell the news" unwind, as institutions booked profits ahead of the technical report and weight release.

The Open-Weight Arms Race

M3's launch is not an isolated event. It lands in the midst of a rapidly intensifying competition among Chinese AI labs to dominate the open-weight segment, with models from DeepSeek (V4), Alibaba's Qwen (3.7), and others racing to claim frontier status .

MiniMax's claim to differentiation rests on being the first and only open-weight model to integrate three capabilities that have historically been delivered separately: frontier-level coding, a 1M-token context window, and native multimodality . The company explicitly frames M3 as designed to match the capabilities of closed-source overseas models, while remaining openly accessible .

This strategic positioning reflects a broader shift in the AI industry's competitive landscape. Where earlier generations of open-weight models competed primarily on cost and speed, the new generation—exemplified by M3—competes directly on capability, pressuring closed-source providers to justify their premium pricing and limited access .

Dual-Listing Ambitions and Capital Strategy

The M3 launch cannot be understood separately from MiniMax's capital-markets strategy. On May 29, 2026, the company signed a sponsorship agreement with CITIC Securities, formally initiating its A-share IPO process targeting the Shanghai Stock Exchange's STAR Market . If successful, MiniMax would become the second Chinese large-model company after Zhipu AI to achieve a dual A+H listing structure .

This follows the company's January 2026 Hong Kong IPO, which raised approximately $619 million (HK$106.7 billion) and delivered a 400% share-price surge by late May . Morgan Stanley had set a price target of HKD 990 on the stock, driven by better-than-expected commercial traction that saw annual recurring revenue jump from $100 million to $150 million in two months .

A dual listing on mainland China's premier technology exchange would deepen MiniMax's access to domestic capital markets while raising its visibility among institutional investors . The simultaneous pursuit of open-weight technical leadership and dual-listing financial firepower reflects a new phase of the Chinese AI competition, where startups must win on both engineering and capital grounds to survive.

M3's launch day encapsulated this tension. A technically ambitious model, backed by benchmarks that demand independent validation, debuted alongside a stock that rose on the morning and fell harder by the close. The next ten days—when the technical report and open weights are scheduled to appear—will test whether the model's benchmarks hold up, and whether the strategic bet on open-weight frontier capability translates into lasting market confidence.