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

MiniMax deliberately shipped M3 as a unified foundation model, not a patchwork of specialized parts. The architecture is natively multimodal, meaning it reads text, images, and video natively, without awkward add-on pipelines. It can even operate a computer desktop directly—a capability that distances M3 from models that treat multimodality as a clunky afterthought .

Notably, MiniMax kept the exact parameter count a secret—a strategic dodge that has become the norm among Chinese frontier labs jockeying for position in the open-weight space .

Stacking Up Against the Titans: GPT-5.5 and Gemini 3.1 Pro

MiniMax hung its credibility on a string of public leaderboard results, claiming M3 flat-out outperforms both GPT-5.5 and Gemini 3.1 Pro on key engineering and agent-driven benchmarks:

• SWE-Bench Pro (software engineering): 59.0%, edging past GPT-5.5 and Gemini 3.1 Pro, and pulling close to Claude Opus 4.7 .
• BrowseComp (autonomous web browsing): 83.5, leaving both GPT-5.5 and Gemini 3.1 Pro in the dust and even nosing ahead of Claude Opus 4.7 .
• SVG-Bench (vector graphics): scored above Claude Opus 4.7 .
• OmniDocBench (multimodal document understanding): reported strong, though specific numbers weren't flashed .
• Terminal-Bench 2.1 (command-line agent tasks): 66.0% .
• Claw-Eval (end-to-end agentic performance): the top score among the tested field .

But here’s the asterisk: many of these numbers were produced on MiniMax’s own turf, using agent-specific scaffolding. As of launch day, no independent labs had verified the claims . The company has promised to release the full model weights and a detailed technical report within roughly ten days—a move that will either cement these benchmarks in stone or invite a reckoning from third-party auditors .

Price Wars: Token Plans and API Access

MiniMax didn’t just compete on benchmarks; it threw down a price-competition gauntlet with a tiered subscription model called the Token Plan :

For developers who prefer to pay by the drink, API pricing starts at around $0.30 per million input tokens (a temporary 50% launch promo; the standard rate is $0.60) and output runs about $1.20 per million tokens (normally $2.40) . Billing is context-aware too: prompts under 512K tokens fall into a standard tier, while those that stretch toward the full 1M-token ceiling hit a higher long-context rate .

With prompt caching optimized, the blended cost can plunge as low as $0.06 per million tokens . This is aggressively cheap—a pricing posture designed to squeeze closed-source rivals that charge a stiff premium for comparable frontier capabilities .

Open Weights, with a Time Buffer

M3 is loudly positioned as an open-weight model, but MiniMax didn’t drop the weights at launch. Instead, the company pledged to publish the full model weights and a detailed technical report on Hugging Face and GitHub within about ten days of the June 1 reveal .

That phrasing—“open-weight” rather than “open-source”—has drawn some developer flak. Critics point out that withholding the training data and original recipes stops short of a genuine open-source release . Still, the model is accessible right now through MiniMax Code, the Token Plans, and the API, meaning developers can integrate it into their workflows while they wait for the self-hosting option .

The Stock Market’s Whiplash Reversal

MiniMax’s shares (ticker 00100.HK) opened 5.24% higher on announcement day, climbing to HKD 884 and briefly kissing an intraday peak of HKD 907.5 . The early euphoria evaporated fast. By the closing bell, the stock hadn’t just surrendered its gains—it had cratered 15.7%, shedding 132 HKD points .

Behind the plunge: heavy short-selling, with about $185.33 million in short interest representing 5.93% of turnover . Analysts pinned the rout on profit-taking by investors who had already ridden an extraordinary wave since MiniMax’s Hong Kong IPO in January 2026, which raised roughly $619 million and delivered a 400% surge by late May .

The M3 launch happened just days after MiniMax’s May 29 bombshell: a sponsorship agreement with CITIC Securities to kick off its A-share IPO process on the Shanghai STAR Market . The double-barreled news—a hot new product plus a secondary listing filing—fueled the morning spike, but it also created the perfect setup for a classic “sell the news” unwind, as institutional investors locked in profits ahead of the forthcoming technical report and weight release.

The Open-Weight Arms Race Intensifies

M3’s arrival isn’t a solo act. It drops into a feverish competition among Chinese AI labs scrambling to claim the open-weight crown, with DeepSeek (V4), Alibaba’s Qwen (3.7), and others sprinting to stay in the frame .

MiniMax’s differentiation play is simple but audacious: it claims to be the first and only open-weight model to fuse three historically separate capabilities—frontier-level coding, a 1M-token context window, and native multimodality—into a single release . The company explicitly frames M3 as designed to match the capabilities of closed-source overseas models while keeping the doors wide open for global developers .

This is a strategic pivot that reflects a broader industry shift. Where earlier open-weight generations competed mainly on cost and speed, the new crop—with M3 as its poster child—is fighting directly on raw capability, forcing closed-source providers to explain why their premium pricing and walled gardens are worth it .

Dual-Listing Ambitions and the Capital Endgame

The M3 launch doesn’t make full sense without the capital-markets backdrop. On May 29, 2026, MiniMax signed a sponsorship deal with CITIC Securities, formally launching its A-share IPO effort targeting the Shanghai Stock Exchange’s STAR Market . If successful, MiniMax would become only the second Chinese large-model company—after rival Zhipu AI—to achieve a dual A+H listing structure .

This follows the January 2026 Hong Kong IPO that raised about $619 million (HK$106.7 billion) and saw the stock price surge 400% by late May, drawing a HKD 990 price target from Morgan Stanley . The bank’s bullish call was driven by better-than-expected commercial traction that saw annual recurring revenue rocket from $100 million to $150 million in just two months .

A dual listing on China’s premier tech board would deepen MiniMax’s mainland capital access and raise its profile with institutional investors . The simultaneous pursuit of open-weight technical leadership and dual-listing financial firepower signals a new, more ruthless phase of the Chinese AI competition, where startups must win the engineering battle and the capital fight to survive.

M3’s launch day crystallized this tension. A technically ambitious model, backed by benchmarks that scream for independent validation, debuted alongside a stock that spiked on the morning and crumpled harder by the close. The next ten days—when the technical report and open weights are scheduled to land—will test whether M3’s benchmarks hold up under external scrutiny, and whether the strategic bet on open-weight frontier capability translates into sustained market confidence or gets written off as another hype cycle casualty.