MiniMax Roadmap Revealed: Small-Parameter Rapid Iteration, API Price Hike Imminent

MiniMax’s next move is now on the table. According to internal information leaked from the Linux.do community, the domestic multimodal large-model company is preparing for the release of the M3 series and Hailuo 3, with clear directions for product strategy and pricing logic.

Small Parameters First, Large Parameters Follow

The release rhythm for the M3 series follows a “small steps, fast run” pattern: a smaller-parameter version will launch in the coming weeks, followed by a larger-parameter variant in the next few months. The core objectives of both models are to improve coding capabilities and agent task execution—two key battlegrounds in today’s large-model competition.

This strategy is unsurprising. Based on M2.5’s 80.2% score on SWE-Bench Verified, MiniMax has already established an advantage in agent capabilities. The M3 series clearly aims to consolidate that position while quickly covering cost-sensitive developer segments with the smaller-parameter version.

The tiered parameter release also aligns with industry trends. OpenAI’s GPT-4o and GPT-4o mini, Anthropic’s Claude Opus and Haiku, and Google’s Gemini Pro and Flash all use models with differing parameter scales to serve different use cases. MiniMax’s move shows confidence in its architecture—the MoE (Mixture of Experts) structure allows them to maintain performance while reducing inference costs, forming the foundation for a tiered product lineup.

Hailuo 3: Ambition to Benchmark Seedance 2.0

On the video generation front, Hailuo 3 is scheduled for release in June 2026—about eight months after the October 2025 launch of Hailuo 2.3. Though slower than the text model cadence, an eight-month generation cycle is considered fast given the cost and complexity of video-generation training.

The key detail is that management has confirmed Hailuo 3 is being built on a “native understanding and generation architecture,” explicitly benchmarking ByteDance’s Seedance 2.0. This is an intriguing statement. Seedance 2.0 is ByteDance’s flagship product in video generation, featuring physical accuracy and motion coherence. By benchmarking it, MiniMax signals that Hailuo 3 will go beyond simple image quality improvements and focus on “understanding real-world physical rules” in video generation.

Technically, this aligns with the frontier direction of AI video generation. Relying solely on diffusion models and compute power is no longer enough—models must grasp physical common sense, spatial relationships, and motion logic. Research like PhyT2V from the University of Pittsburgh and MIT’s World Models explores how to give models “world-model” capabilities. MiniMax’s “native understanding generation architecture” is likely an attempt in this direction.

Hailuo 2.3 has already generated 590 million videos, providing valuable feedback signals for training Hailuo 3. If Hailuo 3 achieves physical accuracy close to Seedance 2.0, its competitiveness in the domestic video generation market will increase significantly.

Confidence Behind API Price Increase: Gross Margin Far Above Industry Average

Pricing strategy is the most noteworthy aspect. Management explicitly stated that the release of the 3-series models will serve as the main lever for increasing API prices across different tiers—this is not a test, it’s a clear move.

Why the confidence to raise prices? Because their profit margins far exceed those of peers. Goldman Sachs’s estimates illustrate it clearly:

• MiniMax text API gross margin: ~40%
• Industry average: 20–25%
• MiniMax multimodal API gross margin: 60–70%

The text API margin is twice that of competitors, and the multimodal margin is extraordinarily high. Behind these numbers lies the inference efficiency advantages of the MoE architecture. At trillion-parameter scale, MoE activates only part of the expert network, making inference costs far lower than dense models. This gives MiniMax the room to maintain price competitiveness while achieving higher profitability—or remain cheaper than rivals even after a price hike.

From a business logic standpoint, this is sound. If M3 and Hailuo 3 show significant capability improvements over their predecessors, a price increase becomes justified. Developers are willing to pay for stronger models, provided the price-to-performance ratio remains attractive. MiniMax’s margin cushion allows them to raise prices while still keeping absolute prices relatively competitive.

However, this also means developers will need to prioritize cost optimization. If your application heavily depends on MiniMax APIs, evaluate how the price adjustment will affect your cost structure, and consider adjusting call strategies or adopting alternative models.

Long-Term Plan for Full Modality Integration

MiniMax’s product matrix already spans four modalities: text, voice, video, and music. The M3 and Hailuo 3 releases represent vertical deepening within each modality, but horizontal integration is the more important trend.

Yan Junjie previously stated that both the M3 and Hailuo 3 models are designed for multimodal fusion scenarios and plan to introduce multimodal integration capabilities in the second half of 2026. This means seamless transformation and cross-modal understanding among text, voice, video, and music—no longer isolated APIs, but unified model capabilities.

This direction holds huge potential. For instance, describing a scene in natural language and having the model directly generate a video with matching background music; or uploading a video for the model to extract dialogs, background sound, and visuals for separate editing and recomposition. If such features materialize, they would drastically lower the barrier to multimodal content creation.

But the technical challenges are serious. Cross-modal alignment, unified representation learning, and multimodal reasoning are still active research problems in academia. How far MiniMax can go will be verified only with real products in late 2026.

Industry Position: From Model Developer to Platform Ecosystem

MiniMax’s positioning is evolving from a pure model developer into a full platform ecosystem. Hailuo AI serves as the consumer-facing traffic entry point, with its 590 million generated videos proving strong market acceptance. Meanwhile, the MiniMax open platform supports nearly ten thousand enterprises and handles hundreds of billions of tokens daily—forming the foundation for enterprise monetization.

The releases of M3 and Hailuo 3 will strengthen this growth flywheel. Stronger models attract more developers; more developers bring more data and feedback, which in turn improve the models. Though API price increases may drive away some price-sensitive users, enterprise clients prioritizing capability will keep buying as long as the value proposition remains solid.

In terms of competition, MiniMax faces global players like OpenAI and Anthropic, domestic giants like ByteDance, Alibaba, and Baidu, and startups such as Zhipu, Moonshot AI, and StepFun. Whether the M3 series maintains its edge in agent capabilities and whether Hailuo 3 can catch up with Seedance 2.0 in video generation will directly determine MiniMax’s position in this fierce battleground.

Impact on Developers

If you currently use or plan to use MiniMax APIs, here are a few key points to note:

1. Cost Planning: The price hike signal is clear. Evaluate its impact on your project budget in advance. If your app is price-sensitive, consider a multi-model strategy—use small-parameter models for simple tasks and large-parameter models for complex ones.
2. Capability Upgrade: M3’s improvements in coding and agent functionality may unlock scenarios previously unreachable or poorly supported. If your app involves code generation, automation workflows, or complex reasoning, keep an eye on M3’s performance.
3. Multimodal Integration: If your product requires cross-modal capabilities (e.g., video dubbing, image-text generation, music visualization), the upcoming multimodal integration version in late 2026 may simplify your tech stack. Until then, you’ll still need to stitch together separate APIs.
4. Alternative Options: Don’t put all your eggs in one basket. MiniMax’s capabilities and pricing are evolving—keep monitoring other models (DeepSeek, GLM, Qwen, Kimi, etc.) to ensure you have a Plan B.

Unverified Information, But Logically Consistent

It’s worth noting that this information comes from community leaks and hasn’t been officially confirmed by MiniMax. However, judging by product cadence, technical direction, and business logic, the plans appear well-reasoned. Rapid iteration of smaller models, video generation benchmarking top products, and leveraging new releases to raise prices—all typical of a mature company.

True confirmation will come once the products launch. If the small-parameter M3 version indeed drops within weeks, that will strongly validate the leak’s credibility. Then, examining actual performance, pricing, and API documentation will clarify how well MiniMax’s strategy plays out.

For developers, getting early insight into this information is always valuable. At the very least, in technical planning and cost forecasting, you’ll know where MiniMax is heading and can make better-informed decisions.

---

References

• MiniMax Rumors - Linux.do – Community leak on M3 and Hailuo 3 release plan and pricing strategy
• From “Full Modality” and “High Quality” to AI Platform Ecosystem – Zhihu – MiniMax multimodal fusion architecture design
• MiniMax - AI Platform | AI Toolkit – Overview of MiniMax product matrix and technical capabilities