What it is in one line
MiniMax's general-purpose AI agent, positioned as "your AI companion": say what you want and it builds a website, a mini-game, a product prototype, or a slide deck — instead of answering you sentence by sentence.
Who built it
Shanghai-based MiniMax, founded in early 2022, CEO Yan Junjie (formerly a vice president at SenseTime). The company listed on the Hong Kong Stock Exchange on January 9, 2026 (00100.HK). It is one of the few Chinese labs that validates its models through a consumer product matrix (Hailuo AI, MiniMax Audio, Xingye/Talkie) before selling into the enterprise.
Note on entries: this ranking entry is the "MiniMax - 通用AI Agent" app (your AI companion), a separate product line from Hailuo AI and Xingye. The three business pillars are: general agent, multimodal platform (Hailuo), and AI companionship (Xingye).
Read: Yan is a veteran of vision and algorithms; MiniMax's method has always been fully-modality in-house models plus self-testing through C-end products. The general agent is their highest-per-user-paying product line and the cleanest part of the IPO story.
What it actually does
- Long-horizon task execution → builds web pages, mini-games, product prototypes, and PPTs; drafts a work plan, confirms with the user, then executes, with modules editable back in the conversation
- Multimodal invocation → the M-series models (M2, M2.1, M2.5, M2.7) let one agent orchestrate text, speech, image, and video generation
- Local execution → MaxClaw, derived from OpenClaw, runs agent tasks on a local machine
- Open platform → enterprise/developer API access, so others can embed agent capability in their own products
The difference from a chat assistant: a chat assistant replies sentence by sentence; this one accepts a goal, breaks it into steps, calls tools, and returns a finished artifact.
What old behavior it replaces
The act of finding a separate tool or person for each deliverable.
To build a website prototype, you used to open a design tool, write code, or hire an outsourcer and a colleague; to build a deck, you opened Office and arranged slides by hand. MiniMax Agent compresses those "switch tool, switch executor" actions into one sentence — one person, one machine, state the need, receive the artifact. Its benchmark is the full flow of one person shipping a product; company executives call it "one person running a company" in interviews.
Read: this replacement only holds when stating a need is cheaper than using the tool. For people who can already use the tool it is a convenience; for people who cannot, it opens a door that used to be closed. MiniMax is betting on the second group.
Business model
Two legs: subscriptions and API. C-end agent subscriptions (average $73 per user), and a metered open-platform API (13.2K cumulative B-end users averaging $6,167 each — enterprise spend far exceeds consumer). Company-wide ARR crossed $150M in February 2026; 2025 revenue was $79.0M, up 158.9% YoY, with roughly 73% from international markets.
Hard numbers
- Traffic-board figure: 3.79M MAU, -1.33% MoM, on the domestic growth, domestic, and global boards
- MiniMax Agent's per-user payment of $73 is the highest across product lines (Hailuo is $56)
- Open platform: 132K B-end users, $6,167 average annual spend each
- Company totals: 212M+ cumulative personal users, 100K+ enterprise customers and developers, 200+ countries
- Listed on HKEX January 9, 2026; ARR crossed $150M by February 2026
Four-way read
| Dimension |
Call |
| Founder-product fit |
Founder is an algorithm veteran; the company builds models through to C-end products itself — strong fit |
| Product insight |
Saw the "chat assistants do not deliver artifacts" gap and made long-horizon execution its own product line |
| Execution quality |
M-series multimodal models plus local agent (MaxClaw); the product matrix proves the engineering ships |
| Timing |
"General agent" is the definitional battle every major player is fighting in 2025-2026; MiniMax has paying data but a smaller user pool than top chat products |
The call
Worth watching. It is one of the few players in the "general agent" lane with real revenue numbers: listed, ARR over $150M, $73 average agent subscription, $6,167 average enterprise spend. These are real-business numbers, not concepts.
Separate two things, though. First, 3.79M MAU at -1.33% MoM — the line is not growing; it is a stable small pool. General agents charge well per user ($73) but reach few people. Second, "general agent" is a definitional contest — Doubao, Kimi, Qwen, and Tencent Yuanbao all use the same word. MiniMax's differentiation is fully-modality in-house models plus global distribution built from companionship products, which is not a moat in consumer chat.
Read: the strongest part of its story is "enterprise users spend $6,167 each" — the real money is B-end, not C-end subscriptions. The consumer agent is brand advertising; enterprise agent capability is the revenue source. Judge it on the B-end, not the MAU.
What to watch next
① Whether 3.79M MAU turns positive MoM — a signal the agent line is entering a growth phase
② Whether the B-end (132K users, $6,167 average) keeps expanding in the earnings reports
③ Adoption of MaxClaw local execution — the bridge from "making PPTs in a chat box" to "actually working on your computer"
What you can take from it
Product logic: build a separate product line for the "chat assistants do not deliver artifacts" gap, and state clearly how it differs from a chat assistant (takes a goal, breaks it down, calls tools, returns a finished artifact) instead of mixing it into the same-named product. Unclear definitions are the common disease of the general-agent lane; whoever states the difference first wins.
Pricing structure: the same line splits consumer subscriptions ($73/user) from B-end API ($6,167/user) — two legs, letting enterprise metering cover costs while consumer tier drives reach. This "consumer for volume, B-end for money" structure is directly copyable.
Verdict
Worth watching. Real revenue, real data, listed. The general-agent pool is small but high quality. Track the B-end growth and local execution, not the flat MAU.