An AI company that also thinks about distribution
During the visit to MiniMax in Shanghai on October 13, 2026, the global team introduced its multimodal strategy through voice and video demonstrations and a discussion of the company’s daily work. The applications covered creating, conversing and producing. Questions from the group moved into internal agents, data, responsibility and the business model.
The connection is converting technical capability into recurring use. A model needs an interface, a problem and a route to the customer. The conversation made it possible to examine these layers together.
Research, product and the founder’s background
MiniMax says it was founded in early 2022. It develops multimodal foundation models and brings them to market through its own products and an open platform for enterprises and developers. Its stated ambition is to advance toward artificial general intelligence. The company’s institutional overview.
Yan Junjie, founder, CEO and CTO, studied mathematics at Southeast University, completed a doctorate in AI at the Chinese Academy of Sciences and conducted postdoctoral research at Tsinghua. Before MiniMax, he spent more than six years at SenseTime. His background helps explain the company’s technical origins. Yan Junjie’s institutional biography.
The company listed in Hong Kong on January 9, 2026. The company’s official filing.
Saving time and spending time are different propositions
The presentation distinguished two intentions: helping someone complete work, and providing companionship or entertainment. Building software, organizing information and producing commercial material belong to the first. Interacting with characters and exploring conversational experiences appeared in the second.
This distinction helps with product thinking. Work can be evaluated through completion, quality and rework. Entertainment also depends on interest, continuity and connection. The same capability serves different needs. Talkie was presented in this context of characters and companionship.
Modalities form a chain of work
Text, voice and video were presented as different ways of using artificial intelligence. In practice, a campaign can start with an idea, move into a script and imagery, acquire motion and finish with narration. The work includes choosing, reviewing and maintaining the intention across these stages.
In the public portfolio, M3 combines coding, agent tasks and image and video inputs, with context of up to one million tokens. The technical introduction to M3.
The product question is where this combination removes a specific difficulty. Many functions together need an understandable workflow and an outcome that meets the person’s objective.
A good voice needs expression, timing and stability
The audio discussion emphasized three criteria: expressiveness, response time and stability. Long narration requires comfort, rhythm and consistency. A conversation needs to respond at the right moment. The team explained why these uses lead to different choices, even when both rely on speech synthesis.
Speech 2.8 is publicly described with features for breathing, hesitation and other details of speech. The official description of voice features.
A voice agent was explained as a sequence: understand the speech, develop a response with the model and context, and turn it into audio. In Brazil, testing should include Portuguese, names, accents and interruptions throughout a task. A pleasant sample does not reveal the complete interaction.
Editing with control changes the proposition of generative video
The video discussion emphasized editing and continuity alongside scene generation. The team gave marketing and gaming examples in which characters, appearance and references need to remain recognizable. The value lies in changing what was requested while preserving what should stay the same.
H3 combines text, image, video and audio references and can generate video with stereo sound. Its weight release carries a specific license; the H3-Context-IR context interpreter remains hosted. H3 architecture and weight release.
This distinction matters in production. Creating something visually interesting is one stage. Delivering a usable asset involves control, revisions, continuity across versions and integration with existing materials. Available weights also do not remove infrastructure requirements or operating costs.
Model, execution environment and context
The team organized AI value into three layers: the model, the environment that allows an agent to work, and a company’s specific context. The term harness was used for the second layer, which brings together tools, execution and the organization of actions.
| Layer | What it adds to the product |
|---|---|
| Model | The ability to interpret, generate and reason about inputs. |
| Execution environment | Tools, integrations and the organization of work stages. |
| Context | Business data, knowledge, criteria and processes. |
A startup needs to identify its contribution: building the experience, understanding a process or connecting specific knowledge. Alongside choosing a supplier, it must organize the relationship between layers and decide who maintains each one.
The company’s own operation becomes a field of use
When the group asked about internal AI use, the team described agents supporting business development. They research companies, compare information with the desired customer profile, investigate further, and help prepare messages and proposals. Meeting notes and CRM records were also part of the explanation.
The most useful aspect of this account is the sequence. Finding a company does not finish the work. Its needs must be understood, fit assessed, the approach prepared and learning preserved for the next conversation. The agent participates in a process that already has objectives and criteria.
A bounded starting point would be an account summary with verifiable sources, leaving the approach decision with the responsible person. Any benefit needs to account for completed work and necessary corrections.
People, data and responsibility remain in the workflow
The recruitment discussion was direct: according to the team, HR still exists and uses AI as support. Finance, legal and marketing were also mentioned as areas with agents. The position defended was to keep decisions and responsibility with people, even when part of the preparation is automated.
Another point was ready access to already digitized data. The representative described permissions that vary by role, with support from IT. This helps explain why context involves preparation, organization and access, rather than just document volume.
Who may consult each piece of information and confirm each action? Suggestions, drafts and execution are different stages. The workflow needs to make them visible to whoever is accountable for the result.
Customer demand can guide technical work
The team described close cooperation between business, marketing and technical teams. Signals from marketing and gaming usage led to discussions about visual coherence, data and product improvement. The example shows how a commercial need can be translated into engineering work.
Feedback needs specificity to become a priority. Losing a character’s identity, requiring many revisions or interrupting a task are observable problems. Benchmarks help compare capabilities; adoption also depends on the customer’s workflow and the cost of correcting results.
The developer is also a distribution channel
The team presented subscriptions and APIs as different routes to market. Its own products deliver a ready-made experience. Integration allows another company to combine the model with its interface, its data and the relationship it maintains with the customer.
Public documentation offers interfaces compatible with Anthropic and OpenAI integration patterns, alongside AI SDK. The API platform access guide.
Less technical friction supports experimentation, but adoption also needs examples, support and evaluation. The conversation covered VCs and accelerators as channels for discovering startups and needs.
Growth needs delivery capacity and workable economics
Demand raised another discussion: training a model and making it reliably available are different kinds of work. The team identified delivery capacity as a challenge and discussed collaboration with third-party infrastructure. For a customer, availability, waiting time and service continuity contribute to perceived quality.
In unaudited first-half 2026 results released on August 26, MiniMax reported US$116.6 million in revenue, up 283.1% year over year. The open platform and other enterprise services represented 63.4% of the total. The period included US$296.9 million in R&D and an adjusted net loss of US$293 million, a non-IFRS measure. Growth therefore coexists with substantial investment and losses. The official half-year results.
For startups, cost per useful outcome matters: generation, review, waiting and rework together. Volume or an impressive demonstration does not establish economic viability.
Localization and rights are part of expansion
The presentation considered differences in language and use across markets. A narration voice and a customer service voice require different evaluations. Expansion involves understanding each audience’s habits, vocabulary, channels and expectations, alongside providing a language option.
An October 2026 company announcement illustrates another dimension: with KAGAMIAI, the proposed H3 IP Edition incorporates approved assets, generation rules, commercial terms and reporting. Access depends on each partner’s conditions. The public intellectual property and AI initiative.
Control and trust contribute to distribution. Working with characters and brands requires defining permitted materials and how outputs enter operations. This organization deserves attention alongside content generation capabilities.
A next pilot for Brazilian founders
The visit suggests starting with a specific problem and testing the combination of model, tools and context. A small pilot observed through to completion can teach more than adding several functions without knowing which one changes the work.
- Choose a task. Define who needs it, the expected result and how they work today.
- Prepare the context. Organize information, criteria and access before automating decisions.
- Evaluate in Portuguese. Include real industry language, corrections and incomplete cases.
- Observe the complete workflow. Measure completion, time, cost and revisions, alongside the first output.
- Preserve control. Separate drafting, approval and execution with clear responsibilities.
- Connect use and improvement. Carry specific customer difficulties into the next version.
Research, product and distribution meet in the combination: turning available capability into a job well done, with an experience people can incorporate into their routine.