From multimodal pioneers to M3.1 Flash and H3
20 October 2026 | BunCosmos Research Lab AI models, creative production and autonomous agents
When people talk about generative video, Hailuo AI often comes up alongside Sora, Kling and Runway. Yet treating Hailuo simply as a video generator misses the larger picture. MiniMax is building across language, video, speech and music, with products that bring these capabilities into creative and agentic workflows.
H3, MiniMax Design and the 19 October release of M3.1 Flash make that direction especially interesting for creators. This deep dive explains how the pieces fit together, what the published specifications support, and where launch claims still need independent testing. The opportunity is a coordinated production workflow, not a guarantee that every generated shot will be ready to publish.
1 Who is MiniMax
MiniMax describes itself as founded in early 2022. Its leadership includes founder Yan Junjie, formerly a vice president and vice-head of SenseTime’s research institute, and co-founder Yeyi Yun. Its consumer products, including Talkie, sit alongside a developer platform and an expanding multimodal model family. [1, 2, 15]
A useful milestone was the January 2025 release of MiniMax-01. Its hybrid architecture combined Lightning Attention with conventional attention and supported context lengths up to four million tokens. That is an important part of the company’s long-context story, but a large context window alone does not establish reliable memory or long-running agent performance. [3]
The naming ecosystem
MiniMax is the company and model provider. Hailuo AI is a creator-facing product. The M-series covers language, reasoning and agentic models, while the Hailuo and H3 families focus on visual and audiovisual generation. MiniMax Design is a creative application that orchestrates models and tools, rather than a new video model. Keeping those layers separate makes the rest of the story much easier to follow.
2 From silent clips to audiovisual H3
Hailuo’s earlier video models developed a reputation for camera movement, physical action and expressive characters. Hailuo 2.3, introduced on 28 October 2025, emphasized motion, stylization and facial micro-expressions. It supported 6- or 10-second video outputs, depending on the selected configuration. These earlier workflows still needed separate audio production. [4, 5]
H3 arrives
MiniMax launched H3 on 31 July 2026. Its core proposition is joint audiovisual generation: text, images, video and audio can inform a scene whose picture and stereo soundtrack are generated together. Dialogue, footsteps, ambience and music no longer have to begin as completely separate tasks. [6]
Documented output and reference capabilities
The published H3 specification lists 4–15-second outputs at 24 frames per second, with 32 kHz stereo audio. The base model generates at 768p short-side resolution; the complete system can produce up to 2K through H3-Regenerate-2K. [7]
H3-Base comes in two main variants. FL2VA supports text-to-audio-video and optional first and last image frames. Ref2VA supports mixed references: up to nine images, three video clips and three audio clips, with a maximum of twelve files in total. Video and audio references each have additional duration limits, and audio cannot be the sole input. [7]
For a creator, this opens an appealing possibility: take the character from an image, movement from a video and sound cues from an audio reference, then describe how they should work together. It is still important to inspect the result. Reference support is not a promise of perfect identity, timing or continuity.
What is happening under the hood
MiniMax describes four major ingredients: contextual multimodal understanding, a redesigned VAE, a unified audiovisual transformer and in-context regeneration. The last is particularly useful to understand: H3-Regenerate-2K feeds a base result and its original context back into the system to recover high-resolution detail, rather than relying only on a generic upscaler. These are the company’s architectural descriptions, not independent measurements of finished-film quality. [6, 7]
3 Open weights and the creative studio
What the public H3 release includes
MiniMax released H3-Base weights on 3 August under its community license. The public release includes FL2VA and Ref2VA, so mixed-reference generation is not inherently cloud-only. The hosted H3-Context-IR preprocessing system and H3-Regenerate-2K are not included in those public weights. The official full workflow can combine a locally deployed base model with hosted components. [7]
That distinction matters. “Open weights” does not mean the complete hosted product is available offline, or that its license has no conditions. Creators planning commercial use should review the applicable model and service terms. Longer storytelling also still calls for shot planning, continuity review and editing beyond one generation.
FastH3 is a separate acceleration effort
FastVideo’s FastH3 Preview v1, announced on 27 August, is a post-trained derivative of MiniMax H3 developed with Nuva Lab and NVIDIA collaborators. The team reported up to 14× acceleration on a single Blackwell GPU and under 13 seconds for a 15-second 768p generation on eight B200 GPUs. These are different test conditions, not a promise of that speed on a laptop or the standard Hailuo service. The initial preview focused on text-to-video-and-audio rather than the complete reference workflow. [8]
MiniMax Design brings the pieces together
Why manually move a brief between a scriptwriter, image generator, video model and audio tool? MiniMax Design addresses that question with a Mac and Windows creative application. Its documented workflow spans planning, script and storyboard work, image and video creation, audio, editing and export, with local asset management and human review checkpoints. [9]
It also reaches beyond MiniMax’s own models. Official product pages list Nano Banana Pro and GPT Image 2 among the image options. Model availability and settings can change, so the current application should be the final reference for a particular production. [10, 11]
Design is an application, H3 is a model family, and a local asset folder does not make hosted generation offline. Generations use credits; check the current plan, credit costs and available models before budgeting a project. Historical launch prices should not be treated as a current quote. [9, 10, 12]
4 M3.1 Flash and agent economics
On 19 October 2026, MiniMax launched M3.1 Flash. If H3 is the studio camera and hands, M3.1 Flash is a candidate for the high-speed brain: a reasoning model that can help interpret materials, plan work and use tools. [13]
Space Bunny unmasked
MiniMax confirmed that Space Bunny, the anonymously tested model, was an early version of M3.1 Flash. The company says that version processed more than 200 trillion tokens. A cosmic rabbit turning up in this story is a very fitting coincidence for BunCosmos. [13]
Context and reasoning
The official API guide documents a one-million-token context window, text, image and video inputs, and reasoning that cannot be disabled. Five effort levels are available, with max as the documented default. These capabilities make long briefs and visual materials relevant inputs, but context capacity is not proof that every detail will be retained or used correctly. [14]
Prices and performance claims
MiniMax’s launch prices are $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. Those are language-model API rates, not the cost of rendering a finished video or buying a Design subscription. [13]
The company reports nearly three times faster online inference and 70% lower usage costs than its predecessor. Its launch chart reports 82.5 on CharXiv RQ. These are vendor-reported results under the stated evaluation conditions. At launch, MiniMax said model weights would be released later. [13]
What this could mean for creative agents
A multimodal reasoning agent could help read a production bible, interpret reference footage, draft a shot list and coordinate generation tools. Paired with a creative application and an audiovisual model, that suggests a practical path from brief to assembled assets. It does not establish that M3.1 Flash is automatically the brain inside every Design workflow.
BunCosmos’s editorial view: the interesting shift is toward directing a production process instead of prompting isolated clips. The test is whether the complete workflow saves effort while preserving visual identity, readable text, usable sound and editorial control. Multi-hour reliability and recovery from mistakes need to be tested, not inferred from token prices or benchmark charts.
5 A quick guide for creators
MiniMax: the company and provider behind the model families and creative products.
Hailuo AI: a creator-facing route to video generation. Check the selected model and account settings for the options actually available.
H3: audiovisual generation with public 768p base weights and separate hosted components for the official context-processing and 2K workflow.
MiniMax Design: the creative workspace that coordinates models and tools. A desktop app and local assets should not be confused with fully local model inference.
M3.1 Flash: a multimodal reasoning model with long context and low published token prices. It can support agent workflows; finished-video cost and quality depend on the whole production process.
Before pressing render
Give each reference a clear job: character, setting, composition, motion or sound. Start with one short scene and judge the result before expanding it into a sequence. Check faces and hands, product details, on-screen text, dialogue timing and transitions. Keep the approved references and successful prompts together so later shots have a consistent starting point.
That is the practical promise behind the agentic studio: more room to direct, with human judgment still responsible for the final cut.
Research and product details checked as of 20 October 2026. Specifications and performance statements above are attributed to their sources.
Primary sources
Numbered references correspond to the linked sources below. Prices, available models and product features can change after the date of this edition.
1 MiniMax investor company profile
2 MiniMax management and cofounder background
3 MiniMax 01 release and architecture
6 H3 launch and design philosophy
7 H3 open weights specifications and hosted components
10 Nano Banana Pro in MiniMax Design
11 MiniMax Design tools and model lineup
12 Current MiniMax Design plans
13 M3.1 Flash launch and benchmark chart