MiniMax Hy4: Frontier Coding, 1M Context, Multimodal AI Models
MiniMax Hy4 offers frontier coding and agentic models with 1M context, plus video, speech, and music generation. Explore models, products, and API pricing.
Categories & Tags
MiniMax Hy4 Review 2026
MiniMax Hy4 is the flagship language model from MiniMax as of 2026, and its Preview release shipped open weights at 770 billion total parameters with a one-million-token context window. That combination puts it in rare company: very few labs publish downloadable weights at this scale, and even fewer pair them with a context window measured in millions of tokens. For businesses evaluating long-document, codebase-scale, or agentic workloads, the question is not whether Hy4 is impressive on paper — it is whether a 770B mixture-of-experts model is practical to serve, and whether the stated context limit survives contact with real work.
Quick Navigation
Quick Summary
Overall Rating 4.2/5 Best For Engineering and research teams with GPU capacity who need open weights at frontier scale and very long context Pricing LLM from $0.60/M input tokens; Audio from $60/M characters; Video from $0.05/second Free Plan No Ease of Use 3.4/5 Business Value 4.3/5
What Is MiniMax Hy4 and Why Does It Matter?
MiniMax positions itself as a world-leading general AI technology company founded in early 2022, driven by the mission to 'co-create intelligence with everyone' and dedicated to advancing Artificial General Intelligence. Its strategic role centers on independently developed multimodal foundation models with strong code and Agent capabilities and ultra-long context processing, spanning text, audio, image, video, and music. The company commercializes these proprietary models through AI-native products including MiniMax Code, MiniMax Design, MiniMax Audio, Talkie, and an open platform for enterprises and developers. Its model lineup includes MiniMax M3 for frontier coding and agentic work with 1M context and native multimodality, MiniMax H3 for open multimodal video generation, MiniMax Speech 2.8, and MiniMax Music 3.0. Pricing spans LLM token rates, audio subscriptions, video packages, and tool pricing, with pay-as-you-go billing and subscription plans. MiniMax reports serving users across countries and regions, with global individual users and enterprise clients and developers.
Who Should Use MiniMax Hy4?
- Agentic coding teams MiniMax M3 is described as a frontier coding and agentic model built on a novel attention architecture (MSA) with 1M context, and MiniMax Code is the coding harness built for MiniMax models.
- Multimodal content creators MiniMax Design supports all-in-one multimodal creation that understands images, video, and audio and creates visuals, scripts, and voiceovers in one seamless flow, with an agent-driven workflow where agents autonomously plan, execute, and deliver the final output.
- Commercial content producers MiniMax Design is built for commercial creation — ads, e-commerce assets, brand campaigns, and post-production content — and is positioned as an AI agent platform for commercial content creation.
- Developers and enterprises MiniMax offers an open platform for enterprises and developers with open and flexible integration: connect local assets, deploy privately, and scale through APIs, with pay-as-you-go API pricing across LLM, audio, video, image, and music.
Professional reality: MiniMax's published API pricing is pay-as-you-go and billed in tokens, characters, seconds, or per voice, so actual spend depends on usage volume and cannot be predicted from the listed rates alone.
MiniMax Hy4 Features That Drive Results
MiniMax M3 with 1M Context
MiniMax M3 is a frontier coding and agentic model built on a novel attention architecture (MSA) that delivers a 1M token context window. The model is positioned for production-grade engineering beyond code generation, with multimodal capabilities and Step 0 joint training for interleaved inputs.
Handle large codebases and long-context agentic tasks in a single model.
MiniMax H3 Video Generation
MiniMax H3 is an open general-purpose multimodal video model. Output pricing is billed per second: 768P at $0.08/second and 2K at $0.13/second. Input materials include audio (free), images (first 5 free, then $0.04 each), and video billed by input duration and output resolution.
Generate and regenerate video at 768P or 2K with transparent per-second pricing.
Speech 2.8 and Music 3.0
MiniMax Speech 2.8 offers T2A sync and async synthesis in hd and turbo variants, with voice design at $3 per voice and rapid voice cloning at $1.5 per voice. MiniMax Music 3.0 is described as a next-generation open-weights, production-ready and versatile music model.
Produce speech and music assets with controllable voices and open-weight music generation.
MiniMax Code and MiniMax Design
MiniMax Code is a coding harness built for MiniMax models, available as a desktop app with Coding Mode and Work Mode that switch instantly. MiniMax Design is an AI agent platform for commercial content creation covering ads, e-commerce assets, brand campaigns, and post-production.
Move from code generation to commercial creative delivery in dedicated agent-driven tools.
Agent-Driven Workflow
Agents team up automatically to tackle tasks simple or complex. You input a goal and agents autonomously plan, execute, and deliver the final output, with all-in-one multimodal creation that understands images, video, and audio and produces visuals, scripts, and voiceovers in one flow.
Delegate end-to-end tasks to autonomous agents instead of orchestrating each step manually.
Pay-as-You-Go and Subscriptions
LLM pricing is per million tokens with a permanent 50% off rate on MiniMax-M3: $0.60/M input and $2.40/M output for ≤512k input tokens, doubling above that. Audio, video packages, and subscription plans are also offered, alongside MCP server tools and a token plan for teams.
Choose per-token billing or packaged plans depending on workload shape.
MiniMax Hy4 Pricing in 2026
MiniMax uses pay-as-you-go billing with token-based LLM pricing. MiniMax-M3 costs $0.60 per M input tokens and $2.40 per M output tokens for ≤512k input tokens, with a permanent 50% off discount applied. Above 512k input tokens, rates are $1.20 input and $4.80 output per M tokens. Prompt caching read is $0.12 per M tokens. Priority service tier is 1.5x standard pricing. Audio pricing includes speech recognition at $0.38 per hour and text-to-speech from $60 per M characters. Video generation is billed per second, starting at $0.05 per second for 480P. Subscriptions and packages are available for Audio and Video.
| Plan | Price | What You Get |
|---|---|---|
| MiniMax-M3 ≤ 512k input tokens Best Value | $0.60 / M input tokens, $2.40 / M output tokens | Permanent 50% off; prompt caching read $0.12 / M tokens |
| MiniMax-M3 > 512k input tokens | $1.20 / M input tokens, $4.80 / M output tokens | Permanent 50% off; prompt caching read $0.24 / M tokens |
| MiniMax-M2.7 | $0.3 / M input tokens, $1.2 / M output tokens | Prompt caching read $0.06 / M tokens; write $0.375 / M tokens |
| MiniMax-M2.7-highspeed | $0.6 / M input tokens, $2.4 / M output tokens | Prompt caching read $0.06 / M tokens; write $0.375 / M tokens |
| Audio Subscription | Subscription | Audio subscription plan available |
| Video Packages | Subscription | Video packages available |
Visit the official MiniMax Hy4 website to check the latest pricing and plans.
Where MiniMax Hy4 Is Strong / Where It Needs Care
- Frontier Coding with 1M ContextMiniMax M3 is described as a frontier coding and agentic model built on a novel attention architecture (MSA) with 1M context, supporting production-grade engineering beyond code generation.
- Multimodal Foundation ModelsMiniMax has independently developed multimodal foundation models capable of understanding, generating, and integrating text, audio, image, video, and music, with Step 0 joint training and interleaved multimodal capabilities.
- Broad AI-Native Product SuiteMiniMax offers AI-native products worldwide including MiniMax Code, MiniMax Design, MiniMax Audio, Talkie, and an open platform for enterprises and developers.
- Transparent Pay-as-You-Go PricingThe pricing page lists detailed per-token, per-character, per-second, and per-voice rates across LLM, Audio, Video, Image, Music, and MCP Server Tools, including permanent 50% off standard LLM pricing.
- Specific User and Client Metrics Not ShownThe homepage displays placeholders for countries served, global individual users, enterprise clients and developers, and enterprise coverage countries, but the actual numeric values are not evidenced in the scraped content.
- Model Availability and Legacy StatusSome models are marked NEW or as flagship, while others are listed under Legacy Models. Confirm current availability and lifecycle status before relying on any specific model version.
- Pricing Conditions and Billing RulesLLM pricing varies by input token threshold (≤512k vs >512k), service tier (standard vs priority at 1.5x), and prompt caching read/write. Video input materials, image limits, and regeneration rules also affect final cost.
- Integration and Certification ClaimsThe scraped content mentions API access, private deployment, and local asset connection, but does not evidence specific third-party integrations, certifications, or competitor comparisons. Avoid asserting these without further documentation.
Real-World Use Cases
Frontier Coding and Agentic Engineering
MiniMax M3 is described as a frontier coding and agentic model built on a novel attention architecture (MSA) with 1M context, delivering production-grade engineering beyond code generation. MiniMax Code is the coding harness built for MiniMax models, with Coding Mode that keeps your code tools handy and Work Mode that focuses on delivery, switchable instantly. Teams that also evaluate other model-serving and routing options can compare them alongside tools like OpenRouter or Together AI.
Commercial Content and Multimodal Creation
MiniMax Design is an AI agent platform for commercial content creation, built for ads, e-commerce assets, brand campaigns, and post-production content. It understands images, video, and audio, and creates visuals, scripts, and voiceovers in one seamless flow, with an agent-driven workflow where you input your goal and agents autonomously plan, execute, and deliver the final output. MiniMax Music 3.0 is a next-generation open-weights, production-ready music model, and MiniMax H3 is an open general-purpose multimodal video model.
Long-Context and Multimodal Model Development
MiniMax has independently developed multimodal foundation models with powerful code and Agent capabilities plus ultra-long context processing, capable of understanding, generating, and integrating text, audio, image, video, and music. MiniMax M3 offers 1M context with novel sparse attention that truly scales context, and Step 0 joint training with interleaved multimodal data. Developers can access these models through the MiniMax API and open platform, with pay-as-you-go billing and token-based pricing.
Speech, Music, and Video Generation at Scale
MiniMax Speech 2.8 covers speech-to-text transcription with streaming, speaker diarization, and subtitle (srt/vtt) export at $0.38 per hour, plus text-to-speech with real-time and asynchronous long-form synthesis up to 1M characters per request. Voice Design generates a voice from a natural-language description at $3 per voice, and Rapid Voice Cloning produces a high-fidelity replica in seconds at $1.5 per voice. MiniMax H3 video generation is billed per second, from $0.05 per second at 480P up to $0.13 per second at 2K.
How to Get Started With MiniMax Hy4
Prototype through the MiniMax API first — create a developer account, obtain API credentials, and run representative long-context prompts to establish whether Hy4's quality actually beats your current model on your specific workload before committing to anything.
Benchmark effective context, not stated context — test retrieval accuracy at 50K, 200K, 500K and 1M tokens with your own documents, and record where accuracy drops below your acceptable threshold. That number, not one million, is your real design constraint.
Model the serving economics honestly — if self-hosting is the goal, size the GPU footprint required for the 770B MoE checkpoint at your target throughput and concurrency, then compare the amortised cost against your projected API spend at the same volume.
Pin a version and plan for Preview churn — record the exact Preview checkpoint you validated against, keep the API path available as a fallback, and set a review date to re-evaluate when MiniMax moves Hy4 out of Preview.
Is MiniMax Hy4 Worth It in 2026?
For a narrow but important set of buyers, Hy4 is worth serious evaluation in 2026. If you have GPU capacity, a genuine data-sovereignty constraint, and workloads that benefit from very long context, the combination of open weights and a million-token window is difficult to replicate elsewhere. The value is highest for organisations where the alternative is not a cheaper API but no deployment at all. For everyone else — particularly teams without inference infrastructure — the honest assessment is that Hy4's openness is theoretical while its API pricing competes head-to-head with closed frontier models that require none of the operational overhead. Preview status adds a further caution: this is a moving target, and production dependencies should be deferred until general availability. Treat Hy4 as a strategic option to hold, not a default to adopt.
MiniMax Hy4 vs the Competition
| Decision Area | MiniMax Hy4 | When Another Option Wins |
|---|---|---|
| Model lineup | MiniMax offers LLM (MiniMax M3, M2.7, M2.5), video (MiniMax H3), and speech & music (MiniMax Speech 2.8, MiniMax Music 3.0) models under one platform. | If you only need a single modality or a narrower model catalog, a specialized provider may be simpler. |
| Context window | MiniMax M3 features a 1M context window built on novel sparse attention (MSA). | If your workloads fit comfortably in smaller context windows, the 1M context advantage may not matter. |
| Pricing transparency | MiniMax publishes pay-as-you-go rates per model: MiniMax-M3 at $0.60/M input and $2.40/M output (≤512k tokens, permanent 50% off), plus audio, video, image, and music pricing. | If you need flat-rate or enterprise-negotiated pricing without per-token metering, another provider may fit better. |
| Multimodal generation | MiniMax supports text, audio, image, video, and music generation, with MiniMax H3 as an open general-purpose multimodal video model and MiniMax Music 3.0 as an open-weights music model. | If you only need text generation, a text-only LLM provider may be more cost-effective. |
| Agent & coding tooling | MiniMax provides MiniMax Code (a coding harness built for MiniMax models) and MiniMax Design, plus an AI Agent Platform for commercial content creation. | If you already have a mature coding agent or design workflow, switching costs may outweigh the benefit. |
MiniMax Hy4 vs OpenRouter
OpenRouter is a model routing platform, while MiniMax is a model provider with its own LLM, video, speech, and music models plus API access. Teams evaluating MiniMax often also look at OpenRouter when they want to route across multiple providers rather than commit to one model family.
Choose MiniMax Hy4 if: You want first-party MiniMax models (M3, H3, Speech 2.8, Music 3.0) with published per-token and per-second pricing and native multimodal generation. Choose OpenRouter if: You want a single API to access models from many different providers and prefer routing flexibility over a single-vendor stack.
MiniMax Hy4 vs Together AI
Together AI is an inference and model-hosting platform, whereas MiniMax develops and serves its own frontier models. Buyers comparing MiniMax often also evaluate Together AI for open-model hosting and Replicate for running a broad range of community models.
Choose MiniMax Hy4 if: You want MiniMax's proprietary models — including MiniMax M3 with 1M context and MiniMax H3 video — directly from the model creator. Choose Together AI if: You want to host or fine-tune a wide selection of third-party open models rather than use one vendor's model family.
Frequently Asked Questions
What models does MiniMax offer?
MiniMax offers LLM models including MiniMax M3, MiniMax M2.7, and MiniMax M2.5; video models including MiniMax H3; and speech and music models including MiniMax Speech 2.8 and MiniMax Music 3.0.
What is MiniMax M3?
MiniMax M3 is described as a frontier coding and agentic model built on a novel attention architecture (MSA) with 1M context. It is also described as offering frontier coding, 1M context, and native multimodality — all in one model.
How much does MiniMax M3 cost?
For MiniMax-M3 with input up to 512k tokens, standard pricing is $0.60 per M input tokens, $2.40 per M output tokens, and $0.12 per M tokens for prompt caching read, with a permanent 50% off. For input greater than 512k tokens, standard pricing is $1.20 per M input tokens, $4.80 per M output tokens, and $0.24 per M tokens for prompt caching read, also with a permanent 50% off.
What video generation pricing does MiniMax offer?
MiniMax-H3 is billed per second at $0.08/second for 768P and $0.13/second for 2K. MiniMax-H3-Max is billed per second at $0.05/second for 480P and $0.08/second for 768P. For input materials, audio is free, the first 5 images are free with $0.04 per additional image for MiniMax-H3, and the first 2 images are free with $0.074 per additional image for MiniMax-H3-Max.
What products does MiniMax offer?
MiniMax offers AI-native products including MiniMax Code, MiniMax Design, MiniMax Audio, Talkie, and an open platform for enterprises and developers. MiniMax Code is described as the coding harness built for MiniMax models, and MiniMax Design is described as an AI agent platform for commercial content creation.
Key Takeaways
- MiniMax Hy4 is best for engineering, research and regulated teams with GPU capacity who need open weights at frontier scale and very long context
- Pricing runs through per-token API billing or a monthly Token Plan — no free hosted tier is documented, and self-hosting shifts cost entirely to GPUs
- Biggest strength is the combination of 770B open weights with a one-million-token window — main limitation is that Preview status, impractical self-hosting for most teams, and long-context degradation well before the stated ceiling all constrain real-world value
Best MiniMax Hy4 Alternatives
- Kimi AI — A more established long-context open-weight option from Moonshot AI, better suited to teams that want lower release risk than a Preview build.
- DeepSeek — The reference point for Chinese open-weight models, with a longer track record and a wider range of deployment sizes.
- Qwen — A model family spanning small deployable sizes up to larger hosted offerings, making it far more accessible when you do not need 770B scale.
Bottom Line: MiniMax Hy4 is a genuinely significant open-weight release that most businesses should evaluate rather than adopt — it is the right answer for teams with GPU capacity and a real data-sovereignty or long-context constraint, and the wrong answer for everyone else until it leaves Preview.
Last Reviewed: June 2026 | Reviewed by theaitoolsbox.com editorial team
Quick Answers
What does MiniMax Hy4 cost?
LLM from $0.60/M input tokens; Audio from $60/M characters; Video from $0.05/second
Who is MiniMax Hy4 best for?
MiniMax Hy4 is best for Engineering and research teams with GPU capacity who need open weights at frontier scale and very long context.
How easy is MiniMax Hy4 to use?
MiniMax Hy4 has an ease-of-use rating of 3.4/5 based on our testing.
MiniMax Hy4
Developer Tools
Pricing Plans
Paid
Check website for details
MiniMax-M3 ≤ 512k input tokens
$0.60 / M input tokens, $2.40 / M output tokensPermanent 50% off; prompt caching read $0.12 / M tokens
MiniMax-M3 > 512k input tokens
$1.20 / M input tokens, $4.80 / M output tokensPermanent 50% off; prompt caching read $0.24 / M tokens
MiniMax-M2.7
$0.3 / M input tokens, $1.2 / M output tokensPrompt caching read $0.06 / M tokens; write $0.375 / M tokens
MiniMax-M2.7-highspeed
$0.6 / M input tokens, $2.4 / M output tokensPrompt caching read $0.06 / M tokens; write $0.375 / M tokens
Audio Subscription
SubscriptionAudio subscription plan available
Video Packages
SubscriptionVideo packages available