MiniMax’s Next Bet: Agents Inside Chips, Labs, and Enterprise Systems

Set the reported $800 million ARR aside for a moment. It is the number that grabs the room, but the more revealing part of MiniMax’s recent call was the plan built around it: M3.1 to get the training and inference base running cleanly, M3 Pro to push toward harder reasoning and agent work, H3.1 to keep joining video, audio, and understanding, and agents aimed at code, cybersecurity, chip design, and drug research.

This is not a plan for a slightly better chatbot. MiniMax wants a model to become an operating part inside a workflow: read the material, use the tools, run the test, notice the mistake, and keep going. The company’s view of the next AI market follows from that idea. Model demand will increasingly come from work being executed, not people taking turns in a chat window.

The plan is ambitious and expensive. MiniMax reported $116.6 million in first-half 2026 revenue, up 283.1% year on year. Open Platform and other AI enterprise services brought in $73.9 million, or 63.4% of revenue, while overseas revenue reached $70.83 million, or 60.8%. Gross margin rose to 17.9%. Adjusted net loss was $293.0 million and R&D spending was roughly $296.9 million. MiniMax is paying today for a seat in a much tougher market tomorrow.

First bet: agents will consume the next wave of inference

Chat usage has a natural ceiling. Someone asks a few questions, rewrites a few paragraphs, and stops. An agent has a different shape. A software task may require it to understand a brief, inspect a repository, use a terminal and browser, write code, run tests, recover from a failure, and try again. Research, support, and data work have the same pattern. The model keeps getting called back until the job is over.

Reporting on the call described this kind of multi-step work as MiniMax’s main source of incremental inference demand. The logic is plain. Once an agent is attached to a ticketing system, codebase, knowledge base, or approval flow, use can move with the business itself rather than with an employee’s appetite for chatting.

More tokens do not make a durable business by themselves. Enterprises already route jobs across models: low-cost systems for easy summaries, stronger models for difficult code, private environments for sensitive data, separate budgets for video. MiniMax has to earn a place in particular workflows by being faster, more reliable, or cheaper once the whole job is counted. Can long context retrieve the right thing? Do tool calls hold up? Can the system stop safely after a mistake? Can a human audit what happened? Those are production questions.

MiniMax’s view of AI: capability grows where it can be checked

A reported management view from the call puts the strategy plainly: if a task can be evaluated precisely and a reinforcement-learning environment can be built around it, a model can improve through repeated attempts. That moves the question beyond whether a model has memorized enough material. It needs an environment that answers back.

Code either compiles and passes tests or it does not. Security issues can be reproduced. Chip work can be evaluated through simulation, coverage, formal verification, and power-performance-area measures. Drug workflows can move through retrieval, prediction, docking, synthesis feasibility, and eventually experiments. Output has somewhere to go. The model can become more dependable through feedback instead of merely sounding convincing in a chat box.

This shifts competition away from pretraining scale alone and toward data, post-training, evaluations, tool environments, and serving. GPUs remain fundamental, but they are one part of the system. The winner is the company that gets more usable capability from a given compute budget and serves it at a cost customers will accept. MiniMax has framed the constraint bluntly: intelligence may keep scaling, but energy and compute do not. In business terms, throughput, caching, scheduling, latency, and token cost eventually show up in gross margin.

M3, M3.1, M3 Pro, and H3.1 belong on the same board

The released M3 is the most concrete part of the plan. Official materials describe native support for text, images, and video; up to one million tokens of context; a mixture-of-experts design with about 428 billion total parameters and roughly 23 billion activated per token; sparse attention; and open weights. For developers, open weights make evaluation, private deployment, and adaptation easier. For MiniMax, they are also a way to bring users toward hosted inference, enterprise support, and tooling.

The M3.1, M3 Pro, and H3.1 references from the call occupy different positions. M3.1 is meant to make the training and inference foundation reusable. M3 Pro is a bet on the upper range of complex reasoning, coding, and agent work; figures near 3T parameters have circulated, but specifications and a release date remain forward-looking. H3.1 would continue the full-multimodal direction. H3 already handles text, images, video, and audio inputs and can generate video with native stereo audio, at up to 2K and 15 seconds. H3.1 has a harder commercial test: move multimodal capability beyond content creation and into work customers will pay more for.

The division of labor is clear. M3.1 is an efficiency wager. M3 Pro is a capability wager. H3.1 is a multimodal-entry wager. Together they reflect a belief that text, images, video, audio, code, and tool outputs will meet inside the same task chain. Video may become part of how a model understands time, operations, and real-world context. It also brings costly inference, content-governance work, and consumer-distribution fees.

Why chips and drug research are next

Chip design and drug research sound distant from one another. Under MiniMax’s logic, they are similar: high-value work, costly errors, and processes that leave feedback behind. Chip design will not be handed to a model for autonomous tape-out. The nearer path is requirements reading, RTL and verification scripts, test cases, bug finding, EDA orchestration, and joint analysis of code, waveforms, and documents. Compilation, simulation, regression tests, and formal verification can keep answering the agent.

Drug work follows the same discipline. Literature and patent retrieval, target hypotheses, molecule screening, synthesis-route suggestions, experiment planning, and clinical-document analysis are closer than a claim that AI will directly discover a drug. Strong virtual results do not remove wet-lab time, biological noise, data rights, or regulation.

In the next six to eighteen months, the evidence that matters is paid pilots moving into production, shorter cycle times, less rework, and credible permission and audit controls. Partnership announcements and benchmark charts are not enough. The question is whether customers will hand over a high-value slice of their process.

Five bets behind the run rate

The first is enterprise APIs and global delivery. Enterprise services made up nearly two-thirds of first-half recognized revenue and overseas revenue was more than 60%. MiniMax is trying to be a service that developers and companies can plug into globally, not just a domestic model vendor.

The second is agent consumption: longer work creates more calls, provided the work finishes and the all-in cost beats the old process. The third is low-cost scaling. MoE, sparse attention, caching, scheduling, and cluster coordination are all attempts to keep a capability gain from being swallowed by the compute bill.

The fourth is M3 Pro. If a larger model produces independently verifiable gains in coding, hard reasoning, and long-running tasks, it could win higher-value customers and strengthen MiniMax’s position. If it is simply larger and more expensive, the market will treat it as another costly release. The fifth is multimodal vertical work. Text, video, and audio together could open chips, drug research, industrial design, and simulation. They can also bring longer sales cycles and more cost.

The $800 million ARR is the capital behind these bets, not the conclusion. The next answer will sit in less glamorous numbers: whether enterprise accounts renew and expand, whether cost per completed task falls, whether M3 Pro ships and survives independent testing, whether multimodal products bring quality revenue, and whether margin and operating cash improve. Big model stories are easy to tell. Building a system that delivers every month is the hard part.

Sources

• MiniMax: First-half 2026 financial results
• PR Newswire: MiniMax first-half results release
• MiniMax M3 product page, M3 research release, and M3 GitHub repository
• MiniMax H3 release
• Reuters: MiniMax first-half results coverage

Financial and released-product facts are current to August 31, 2026. ARR, internal efficiency, M3.1, M3 Pro, H3.1, and vertical-industry expansion from the telephone call are treated as management statements or public reporting, not as delivered products or recognized revenue.