Key Facts and Timeline

MiniMax released its half-year 2026 financial results on August 26, 2026, marking a validation phase in its commercialization journey. Key hard facts:

• Report Date: August 26, 2026 (H1 results) with August 2026 ARR update
• Revenue (H1 2026): ~$120 million, up 283% year-on-year; half-year revenue equals 1.5x full-year 2025 revenue
• ARR (Annual Recurring Revenue): Over $800 million as of August 2026
• Revenue Structure Shift: B-end (To B) contributes ~80% of ARR; C-end (To C) down to ~20%
• Growth Acceleration: July token consumption reached 20x January levels; enterprise customers and developers exceeded 2 million (10x 2025 year-end)

ARR represents an annualized estimate based on current recurring revenue levels—not recognized revenue in the formal financial statements—but its rapid growth reflects significantly improved commercialization efficiency.

Revenue Structure Reversal: B-End Becomes Primary Growth Engine

MiniMax has undergone a fundamental shift in revenue composition within one year. In H1 2026, open platform and other AI enterprise service revenue reached ~$74 million, up 703% year-on-year; AI-native product (C-end) revenue was ~$43 million, up 101% year-on-year. Enterprise service revenue accounted for 63% of total revenue, up from 30% in the same period last year.

Unexpected contrast: B-end revenue growth (703%) significantly outpaced C-end (101%), completely reversing the proportion landscape—from 30% B-end / 70% C-end a year ago to ~80% B-end / ~20% C-end in August 2026 ARR. MiniMax VP Xue Zizhao stated at the earnings call that third-quarter revenues exhibit acceleration.

Growth drivers include three factors: expanding enterprise customer base, increased API call volumes, and wider adoption of the Token plan. The company now counts over 2 million enterprise customers and developers—10x the 2025 year-end number. Geographically, over 60% of H1 2026 revenue came from overseas markets, with domestic revenue accounting for ~40%.

Agent-Driven Token Consumption Explosion

Rising ARR is primarily driven by per-customer token consumption surges. MiniMax management noted that model consumption has shifted from “human-AI interaction” to “agent-AI interaction”—a single human task request gets decomposed by agents into multiple model requests, tool calls, and sub-agent tasks, causing token consumption to grow faster than user counts or message volumes.

Critical statistic: July token consumption reached 20x January levels. Whether this growth sustainably translates to revenue and margin depends on model pricing, inference cost control, and cluster utilization rates.

MiniMax stated that over the past two months, text model unit compute throughput has increased threefold. M3.1 aims to reduce inference costs to one-third of M3’s initial launch levels. The company emphasizes: “We don’t see pricing Reduction and gross margin improvement as contradictory”—as long as unit token margin remains positive and continuously improving, scale expansion delivers higher absolute gross profit.

Financials show H1 2026 gross profit of ~$21 million (up 465% year-on-year), while R&D expenses remained high at ~$300 million (up 139%), with IFRS net loss of ~$360 million (narrowing from ~$400 million year ago). This confirms revenue structure optimization is underway, yet revenue still falls short of covering model R&D and compute costs.

Product Pipeline Synergy: M3 and H3 Dual Drivers

MiniMax’s B-end call growth is driven by two technical pipelines: language models (M series) and multimodal models (H series).

• M3 Series: Primarily serves Coding, Agent, and long-horizon tasks; core API call model for enterprise clients
• H3 Series: Targets video generation and multimodal content production; open-sourced versions downloaded over 24 million times in three weeks, spawning 300+ public derivative models

MiniMax explicitly positions open source as a strategy to enter enterprise workflows and developer ecosystems—not as a substitute for commercial services. Management stresses: “Long-term pricing power ultimately depends on effectiveness, cost, stability, and iteration speed—not on whether a model is open source.” Large-scale enterprise applications still require efficient inference, stable service, continuous upgrades, and enterprise-grade delivery capability.

Upcoming iterations include M3.1, M3 Pro, and H3.1. M3 Pro is expected to reach ~3 trillion parameters, with expanded reinforcement learning and long-horizon training. Existing and planned compute resources can support continued iteration of 3T-level text and video models. M3 and H3 are also undergoing domestic chip compatibility testing, with large-scale domestic compute clusters gradually taking on real production traffic.

Adoption Recommendations and Market Outlook

Ideal for immediate adoption by:

• Enterprise app developers requiring high-frequency API calls, especially those already using agent frameworks
• B-end customers needing integrated multimodal capabilities (language + vision + audio)
• Cost-sensitive enterprises seeking to benefit from continuously declining unit token costs

Groups advised to wait:

• Enterprises with extremely high stability requirements尚未验证国产芯片适配版本 in production environments
• Pure C-end application teams, as MiniMax resources are clearly shifting toward B-end

Final Thoughts

MiniMax has completed a strategic pivot from “consumer product dominant” to “enterprise service driven” in one year. The $800 million ARR validates the feasibility of large-model commercialization paths. The next critical test will be whether the company can build a sustainable profitability flywheel balancing token consumption surges with falling inference costs.