Mixture-of-Experts Makes Frontier Scale Servable

Hy4 uses a mixture-of-experts design in which only about 49 billion of the 770 billion total parameters are active per token. That sparsity is the entire reason a model of this size can be served at all — a dense 770B model would demand roughly sixteen times the compute per token. For buyers, this means the hardware floor is lower than the parameter count suggests, though still far above what a single workstation can handle.

Business outcome: frontier-class output quality at a fraction of the per-token compute a dense model of the same size would require.

One-Million-Token Window For Whole-Corpus Reasoning

The stated one-million-token context window is the headline capability. In practice it enables workflows that were previously impossible without retrieval scaffolding: loading an entire mid-sized codebase, a full year of board minutes, or a complete regulatory filing set into a single prompt. The caveat is that stated limits and effective limits are different numbers — see the limitations section for how this plays out.

Business outcome: eliminates the engineering cost of building and maintaining retrieval pipelines for document sets that fit within the window.

Downloadable Weights Remove Vendor Lock-In

MiniMax published Hy4 Preview weights on 28 August 2026, placing it among the Chinese labs now shipping open downloads alongside DeepSeek, GLM, Kimi K3 and the smaller Qwen and Ernie models. For a business, downloadable weights mean a model version can be pinned, audited, fine-tuned and deployed on private hardware indefinitely — none of which is possible with a closed API.

Business outcome: protects long-lived products from provider deprecation and removes the need to re-qualify a new model version every few months.

Part Of A Broader Multimodal Model Family

Hy4 sits inside MiniMax's wider model lineup, which spans language, video generation, speech and music — the company describes itself as having independently developed multimodal foundation models capable of understanding and generating text, audio, image, video and music. Teams already using MiniMax's video or speech models can consolidate vendor relationships and, in some cases, share infrastructure.

Business outcome: fewer vendor relationships to manage and a single procurement and security review covering multiple modalities.

Two Delivery Paths — Self-Host Or API

MiniMax offers an open platform for enterprises and developers with API access, alongside the downloadable weights. That dual path lets a business prototype cheaply on the API and migrate to self-hosting once volume justifies the capital expense, or run a hybrid where sensitive workloads stay on-premise and the rest goes through the API.

Business outcome: allows a staged migration from rented to owned inference without rewriting the application layer.

Backed By An Established Commercial Platform

MiniMax is not a research project — it operates AI-native products including MiniMax Code, MiniMax Design, MiniMax Audio and Talkie, plus an enterprise developer platform, and serves users and enterprise clients across a large number of countries. That commercial base means the model is unlikely to be abandoned, and there is a real support and documentation surface behind it.

Business outcome: lower abandonment risk than a purely academic open-weight release, with a documented API and enterprise contact path.

• Open weights at a scale almost nobody else publishesA 770B checkpoint with downloadable weights puts MiniMax in a very small group of labs, and gives buyers a genuine exit from API dependency.
• Sparse activation keeps serving theoretically viableActivating roughly 49B of 770B parameters per token is what separates a servable model from a research artifact.
• One-million-token context removes retrieval engineering for large corporaFor document sets that genuinely fit, the window eliminates an entire class of pipeline work.
• Commercial backing reduces abandonment riskMiniMax runs a real product and enterprise business, so the model is unlikely to be orphaned the way academic releases often are.

• Preview status means the target is movingHy4 Preview is a pre-final release, so behaviour, weights and pricing can change before general availability — do not build a production dependency on Preview behaviour without a migration plan.
• 770B open weights are out of reach for most self-hostersEven with sparse activation, serving this model requires a multi-GPU deployment that the overwhelming majority of businesses cannot justify or operate.
• Long-context performance degrades well before the stated limitA one-million-token window is a maximum, not a guarantee — retrieval accuracy across a full million tokens typically falls off long before the ceiling, so effective context is usually a fraction of the headline number.
• Professional RealityIf your team does not already run multi-GPU inference infrastructure, the 770B open weights are effectively unusable — the API is the only realistic path, and at that point you are comparing Hy4 to every other hosted frontier model on price and latency, not on openness.

Whole-codebase reasoning for platform engineering teams

A mid-sized monorepo often exceeds what standard context windows can hold, forcing teams to build chunking and retrieval layers that lose cross-file relationships. Hy4's window lets an engineering team load substantially more of a repository at once, which improves refactoring suggestions, dependency tracing and migration planning. Teams already using AI coding tools should read our roundup of the best AI coding tools in 2026 to see where a long-context model fits alongside editor-integrated assistants.

Contract and filing review in legal and compliance

Reviewing a deal room or a regulatory filing set means holding hundreds of pages of interdependent language in view simultaneously. A long-context model reduces the risk of missing a cross-referenced clause that a chunked pipeline would separate. For firms weighing this against purpose-built legal tooling, our guide to the best AI legal tools in 2026 covers the specialist alternatives.

Long-horizon agentic workflows

Agent systems that run for many steps accumulate context fast — tool outputs, intermediate reasoning, retrieved documents. A model with a large window can sustain longer autonomous runs before needing summarisation or external memory, which reduces the failure modes that come from aggressive context compression. This is the workload where the combination of open weights and long context is most defensible.

Sovereign deployment for regulated data

Banks, hospitals and government contractors frequently cannot send sensitive text to a third-party API regardless of how good the model is. Downloadable weights let these organisations run frontier-class inference inside their own security perimeter, which is often the only path to approval. The trade-off is that they own the uptime, the GPU bill and the model maintenance.

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.