MiniMax

Key Features

• A Genuine Multimodal Stack: MiniMax does not just offer a text model with a vision add-on. Its model suite spans five fully independent modalities: the “M-series” for language and reasoning, “Speech” for voice synthesis and cloning, “Hailuo” for video generation, and “Music” for AI-generated audio. Each model is production-grade in its own right, and all are accessible through a single unified API. For developers building products that combine text, voice and video, MiniMax removes the need to stitch together five different vendor APIs.
• Sparse MoE Architecture Built for Price-Performance: The flagship M2.7 is a 229-billion-parameter Mixture-of-Experts model that activates only around 10 billion parameters per inference, roughly 4.3% of its total capacity. This architecture is the core reason MiniMax can offer competitive frontier-quality output at API prices that are dramatically below Western equivalents. The intelligence stays high while the compute cost does not.
• Agentic Performance That Scales: MiniMax-M2.5, released in February 2026, was trained with reinforcement learning across hundreds of thousands of complex real-world environments. It scored 80.2% on SWE-Bench Verified and completed that same benchmark 37% faster than its predecessor M2.1, matching the speed of Claude Opus 4.6. M2.5 is specifically engineered for multi-step tool use, browser automation, code interpreter coordination and MCP integrations: the kind of long-chain agentic work that breaks lesser models after a few dozen turns.
• Ultra-Long Context: MiniMax-Text-01 introduced a 4-million-token context window, approximately 31 times the size of GPT-4o’s at launch. It’s capable of ingesting roughly 3 million words in a single prompt. While the M2 series operates at a more practical 205K context window, the underlying engineering philosophy is clear: MiniMax builds for developers who need to process entire codebases, legal archives, or financial document sets in one shot.
• Voice Synthesis and Cloning at Scale: Speech 2.8 can generate natural synthetic voices in 17+ languages, clone a voice from as little as 10 seconds of audio, and adjust cadence, tone, and delivery style dynamically. At roughly $60 per million characters of input for the Turbo tier, one million characters translates to approximately 11–12 hours of spoken audio output making MiniMax one of the most cost-effective professional-grade TTS options available anywhere.
• Hailuo AI Video Generation: Hailuo 2.3 and its Fast variant are MiniMax’s text-to-video models, with significant improvements in character motion, visual quality, and stylistic expression over prior generations. The Hailuo 2.3 Fast variant can reduce batch content creation costs by up to 50%. Generating a 6-second 1080p video via the API costs around $0.33 — a price point that makes it viable for production pipelines rather than just experimentation.

Company Background

MiniMax was founded in December 2021 by Junjie Yan, a former Vice President at SenseTime, along with a team of computer vision researchers from the same firm. The company’s name is derived from the minimax algorithm (a decision-making concept from game theory) and that origin says something deliberate about the company’s design philosophy: optimise for outcomes across competing constraints, not just raw capability.

Early funding came from MiHoYo, the gaming company behind Genshin Impact, with subsequent backing from Alibaba, Tencent, Hillhouse Investment, HongShan, and IDG Capital. By March 2024, an Alibaba-led round valued the company at $2.5 billion. On 9 January 2026, MiniMax listed on the Hong Kong Stock Exchange (HKEX: 00100), with shares roughly doubling on day one to imply a market capitalisation near HK$345 billion making it one of the most closely watched AI listings of the year.

The company’s 2025 annual results showed revenue of $79 million, up 158.9% year over year, with more than 70% of that revenue derived from international markets. While the business remains in investment mode with an adjusted net loss of $250.9 million. Its gross profit margin expanded by 13.2 percentage points to 25.4%, signalling improving unit economics as model efficiency improves. The M2 series, in particular, went from launch to over six times its December 2025 daily token consumption by February 2026, driven primarily by developer adoption in coding pipelines.

MiniMax also made headlines in early 2026 when Anthropic publicly alleged that MiniMax, along with two other Chinese AI companies, used thousands of fraudulent accounts to generate over 16 million interactions with Claude for the purpose of model distillation. MiniMax disputed the characterisation. The allegation has not been adjudicated, but it underscores just how fiercely competitive the frontier model race has become.

User Experience

• MiniMax Agent (The Workspace Product): Launched in January 2026, MiniMax Agent is the company’s answer to the agentic workspace trend. It supports full-modality task execution including text generation, coding, web search, data analysis, and one-click video generation from a simple description. Internally, MiniMax reports that Agent interns now support nearly 90% of employees across software development, data analysis, operations and sales. The platform feels less like a chatbot and more like a task executor.
• Talkie (The Consumer App): Talkie is MiniMax’s AI character and companion chat application, targeted primarily at international consumer audiences. With 11 million monthly active users by mid-2024 and an annualised revenue run rate of approximately $70 million, it is one of the most commercially successful AI companion apps globally while the majority of its users are based in the United States. It runs on M2-Her, a version of the M2 model specifically fine-tuned for long-form, personalised conversational experiences; M2-Her ranked first globally in 100-turn long-context dialogue testing.
• Hailuo AI (The Video Platform): Hailuo AI is MiniMax’s standalone video product. It’s a consumer-facing interface for text-to-video and media agent workflows. It supports full-modality content creation and is broadly available to international users through both web and API access.
• The Developer Experience: For engineers, MiniMax’s Open API Platform is the primary entry point. API documentation is available in English, the SDK is straightforward, and the pricing structure is transparent. The M2 and M2.5 models are also deployable via Claude Code, Cursor, Cline and Kilo Code meaning developers already embedded in those ecosystems can drop MiniMax into existing workflows without rebuilding around a new interface.
• Known Friction Points: MiniMax’s consumer apps and models operate under Chinese content regulations, meaning politically sensitive topics trigger hard refusals with no configuration. Data is hosted in China by default. Enterprise compliance teams in regulated US and EU sectors should evaluate the private deployment option before committing. English-language community support is thinner than Western alternatives; API documentation exists in English, but debugging forums and tutorials lean heavily Chinese. Commercial licensing for M2.7 is also tighter than prior releases and requires written authorisation from MiniMax for any paid product deployment.

Cost

MiniMax’s pricing strategy rewards high-volume API consumption and cost-conscious developers above all else. Consumer apps carry free tiers, and the open-weight models are downloadable at no licensing cost.

Free Tier

• MiniMax Agent (Free Plan) – Available to all registered users.
  • Price: $0
  • Includes: Access to the Agent workspace, AI search, basic coding assistance and task automation with daily credit limits.
  • Use case: Everyday productivity, light coding, document workflows.
• Talkie – Free to download on iOS and Android.
  • Price: $0 base, with in-app purchases for premium character features.
  • Use case: AI companions and character-based interactions.
• Open-Weight Self-Hosting – M2 and M2.7 weights are available for download from Hugging Face.
  • Price: $0 in licensing fees; compute costs apply.
  • Hardware: M2.7 (229B MoE) requires a minimum 128GB Mac or equivalent VRAM configuration for the GGUF quantised version (~108GB); expect 15+ tokens/second on compatible hardware.
  • Use case: Data-sovereign deployments, IP-sensitive workloads, regulated industries requiring on-premises inference.

MiniMax Agent Subscription (Billed Monthly or Yearly)

MiniMax Agent offers tiered subscription plans for users who need consistent access beyond the free daily credits.

• Free – Entry-level access with limited daily credits.
  • Price: $0/month
• Basic – For casual and moderate use.
  • Price: ~$9.99/month
  • Includes higher credit allocation for agentic tasks, coding, and media creation.
• Pro – For active daily use across coding and creative workflows.
  • Price: ~$29.99/month
  • Includes priority access to M2.5, faster response queuing, and increased credit pools.
• Ultra – Maximum credits with access to all modalities.
  • Price: ~$69.99/month
  • Includes priority routing, higher concurrency, and dedicated support access.
• Enterprise – Custom pricing for teams and organisations.
  • Includes SLA-backed support, higher API rate limits, and private deployment options.

Pay-Per-Token API Pricing

For developers who prefer usage-based billing, MiniMax’s token economics are among the most competitive available at the frontier tier.

• MiniMax M2 / M2.5: ~$0.15–$0.30 input / ~$0.95–$1.20 output per million tokens, roughly 8% of the cost of Claude Sonnet at comparable performance tiers.
• MiniMax M2.7 (flagship): ~$0.30 input / $1.20 output per million tokens, approximately 10x cheaper than Claude Sonnet on input.
• Speech 2.8 (Turbo): ~$60 per million characters of input (~$0.000060/char)
• Speech 2.8 (HD): ~$100 per million characters of input
• Hailuo 2.3 Video: ~$0.33 per 6-second 1080p video clip
• Image Generation: ~$0.0035 per image

For full details, see MiniMax API Pricing.

---

In summary, MiniMax is not just a Chinese counterpart to the Western AI giants. It is a multimodal infrastructure play with a product footprint that Western labs have not matched in breadth. It combines frontier-class language models, production-grade voice synthesis, commercially viable video generation and AI-native workspace tooling under a single API and a single company. By pricing aggressively, open-sourcing its weights, and building products that are already generating meaningful international revenue, MiniMax has positioned itself as the platform of choice for cost-conscious developers, multimodal product builders and enterprises who need more than a chatbot. For anyone building AI-native products in 2026, it belongs in the evaluation shortlist.