MiniMax Posts First-Ever LLM Company Earnings Report, Foresees Three 2026 Super-PMFs as AI Platform Play Takes Shape

The Chinese LLM pioneer's IPO debut reveals a path to sustainable unit economics—$79M revenue, 158% growth, and gross margin climbing from 12% to 25%

Just 52 days after its Hong Kong IPO, MiniMax has dropped something unprecedented in the LLM industry: an actual earnings report with real numbers.

In February 2026, the company's Annual Recurring Revenue (ARR) exceeded $150 million. For 2025, total revenue grew 158.9% year-over-year, gross profit skyrocketed 437%, and net loss margin narrowed significantly.

But the bigger story is this: as the first-ever LLM company to publish earnings, MiniMax has opened a valuable window into "how large models commercialize"—and established a key benchmark for measuring whether Chinese AI companies can compete globally.

From this starting point, we can glimpse not just MiniMax's next moves, but the evolution of the entire AI large model industry.

Global First: How Did the Numbers Look?

2025 marked MiniMax's second year of true commercialization, and the company delivered:

Total revenue of $79.04 million, up 158.9% year-over-year, with over 70% coming from international markets.

Adjusted net loss was $250 million, but the net loss margin has narrowed substantially. Simply put: they're making more money and losing less.

This doubling of revenue comes from a dual-engine strategy that's becoming clearer by the day.

Revenue Breakdown: Consumer vs. Enterprise

MiniMax's revenue splits into two main categories:

Consumer AI Products (67.2% of revenue, $53.08M, up 143% YoY): Subscription revenue from MiniMax, MiniMax Voice, Hailuo AI, and Xingye apps. By end of 2025, these apps served 236 million users across 200+ countries—a user base that can compete with top global internet products.

Enterprise Platform & B2B (32.8% of revenue, $25.96M, up 197.8% YoY): Usage-based billing for enterprises and developers. The company now has 214,000 enterprise customers and developers across 100+ countries, with over 50% of platform revenue from overseas.

This "Consumer + Enterprise" dual-engine model provides stable, predictable recurring revenue—and significantly improved profitability.

The Path to Profitability

2025 gross profit hit $20.08 million, up 437% YoY, outpacing revenue growth.

Gross margin improved dramatically: from -24.7% in 2023, to 12.2% in 2024, to 25.4% in 2025.

This improvement came from model and system efficiency gains, plus infrastructure optimization. Combined with narrowing net losses, MiniMax is approaching sustainable commercialization.

R&D Efficiency: Doing More with Less

2025 R&D spending was $250 million (up 33.8%), but as a percentage of revenue, it dropped from 619% in 2024 to 320% in 2025—showing improving efficiency.

Cash reserves stand at $1.05 billion (including cash equivalents and restricted cash), up from $880 million at end of 2024. Post-IPO, rising stock prices have only strengthened this war chest.

Product Velocity: Three Models in 108 Days

2025 was a year of technical acceleration. MiniMax built full multimodal capabilities—language, video, voice, and music models competitive globally.

In the 108 days from Q4 2025 to early 2026, the company shipped three generations of language models:

October: M2 — Returned to full attention mechanisms for real-world stability, "breaking the impossible triangle of intelligence, speed, and cost." This model won first place in the University of Hong Kong's AI-Trader simulated stock competition.

December: M2.1 — Achieved SOTA on multilingual programming and fixed a major Vibe Coding weakness around frontend/backend standards. The OpenClaw author later called it "the best open-source model at the time."

January 2026: M2.5 — Lighter, faster, cheaper. 100 TPS inference speed, full-stack development capability, and just $1/hour as an agent.

On the day of this report, Notion co-founder Akshay Kothari announced M2.5 as the first open-weights model in Notion Custom Agents—a significant enterprise validation.

Agent Platform Evolution

While models advanced, MiniMax also updated its Agent platform twice:

January 2026: Agent 2.0 — Desktop version launched, deeply integrated with local work environments, capable of reading local files directly.

Experts System — Users can create domain-specific experts with unique knowledge and workflows.

Post-CNY Update — Community-driven Experts system strengthened, with over 10,000 applications created.

The platform also introduced MaxClaw mode, enabling one-click cloud deployment directly from the web interface—no complex configuration needed.

Three "Super-PMFs" for 2026

On the earnings call, CEO Yan Junjie outlined three product-market fits MiniMax is preparing for:

1. Programming: L4 to L5 Intelligence

AI coding assistants will evolve from "tools" to "colleague-level collaborators." L4 means innovative solutions to complex engineering problems; L5 requires multiple agents collaborating like human teams.

2. Workplace: The Next Frontier

Every profession's workspace tools represent a larger market than coding. MiniMax believes office scenarios will replicate coding's 2025 progress trajectory.

3. Multimodal Creation: Deliverable-Grade Output

AI creation tools will generate "delivery-ready" mid-to-long-form content, possibly with streaming, real-time output. MiniMax has already upgraded Hailuo AI's Media Agent for full-modal creation and one-click video generation.

From Model Company to Platform Company

The winners of the next generation won't be pure technology providers, but organizations strong across technology, product, and commercialization.

"You must be a doer, not a test-taker," Yan emphasizes. The competition is no longer about benchmark scores, but who can deliver results in chaotic real-world business environments.

MiniMax is explicit about its evolution: transitioning from a "large model company" to an "AI platform company."

The formula: Platform Value = Intelligence Density × Token Throughput

When both factors are strong, platform value becomes inevitable. The upcoming M3 and Hailuo 3 series models are designed specifically for this—optimizing inference architecture and compute efficiency.

2026, in MiniMax's view, is the watershed moment when AI transitions from the "tool era" to the "ecosystem era."