MiniMax M3 Guide: Features, Pricing, Best Use Cases

Most models added vision as a bolt-on. MiniMax M3 was trained multimodal from step zero — text, images, and video sharing one semantic space from the first gradient update — and it shows in how naturally the model moves between a codebase, a screenshot, and a screen recording in a single conversation. Launched at the end of May 2026, MiniMax M3 is one of the most quietly capable releases of the year: an open-weight, reasoning-capable, natively multimodal model priced at $0.30 per million input tokens. This MiniMax M3 guide covers what it is, what it costs, and the use cases where it beats models three times its price.

What Is MiniMax M3?

MiniMax, the Shanghai lab best known until recently for its consumer AI products, announced M3 on June 1, 2026 — though the model reached hosted catalogs a day earlier, on May 31, which is the release date you will see on CoreAI's model list. Open weights followed on Hugging Face within the week. The company positions it as its first release to combine reasoning and agent capabilities in a single open-weight model, per the official MiniMax announcement.

Under the hood it is a mixture-of-experts transformer: 428 billion total parameters with roughly 23 billion active per token, which keeps inference costs closer to a mid-sized model than a frontier one. Long context is served through MiniMax Sparse Attention (MSA), the lab's efficiency technique for keeping attention costs sane at extreme lengths. One note on the numbers: MiniMax's own materials cite a 1M-token window, while the model as served on CoreAI exposes 524,288 tokens of context — still enough for several novels' worth of input, and paired with a 512,000-token output cap that is among the largest of any model in the catalog.

The differentiator is the modality story. M3 was trained on mixed modalities from step zero rather than having a vision encoder grafted on later, so text, images, and video genuinely share one representation space. Video input remains rare even among premium models; in the sub-dollar price class it is nearly unique.

How Much Does MiniMax M3 Cost?

Here is the pricing picture, with per-million-token rates from CoreAI's live catalog for the models a prospective M3 user would shortlist:

At $0.30 in and $1.20 out, M3 sits at the bottom of the standard tier on price while carrying the widest input-modality support in the table. Cached input drops to $0.06 per million tokens, which agentic loops exploit heavily. The practical math: a long agent session consuming 3 million input tokens (much of it cached) and producing 300,000 output tokens lands under a dollar. On Claude Sonnet 5, the same session costs several times that. Qwen3.7 Plus is the closest price rival — our Qwen3.7 family guide covers that lineup — but it does not take video.

What Are the Best Use Cases for MiniMax M3?

Video and screen-recording analysis. This is the headline act. Feed it a screen recording of a bug reproduction and ask what went wrong; summarize a recorded meeting or lecture; extract steps from a tutorial video. Most workflows like this previously required a premium multimodal model or a clunky frame-extraction pipeline. M3 does it natively at budget-adjacent prices.

Agentic coding with visual context. MiniMax reports 59.0% on SWE-Bench Pro — vendor-reported, as always, but in the neighborhood of dedicated coding models. The interesting combination is code plus perception: an agent that reads the failing UI screenshot and the component source in one context. The 512,000-token output cap also means it can emit an entire generated project in one response instead of twenty.

Long-document work with mixed media. Half a million tokens of context holds a contract stack, a slide deck exported as images, and your notes simultaneously. Reasoning support means you can toggle deeper thinking for analysis-heavy questions — see our thinking-mode guide for when that toggle actually pays off.

High-volume production pipelines. Open weights plus low hosted pricing is a flexible combination: prototype on the hosted model, then self-host the identical weights if scale or data policy demands it later.

Where would I pick something else? Pure-text frontier reasoning at the highest difficulty still belongs to premium models, and if your work needs a full million tokens of context in the hosted version, GLM 5.2 or Qwen3.7 Plus serve larger windows. For everything in between, M3's price-to-capability ratio is hard to argue with.

Where does M3 sit among 2026's open-weight releases?

The first half of 2026 has been a stampede of capable open-weight models out of Chinese labs — GLM 5.2 with its million-token window, the Qwen3.7 line, Kimi's coding specialists — and M3 carves out a distinct corner of that map. It is not the biggest window, not the absolute cheapest, and not the top coding benchmark. What it is: the only one in its price band that watches video, reasons, and codes in the same model. That generalist-with-eyes profile is why agent builders have been testing it as an orchestration workhorse — one model that can read the ticket, watch the reproduction clip, patch the code, and write the summary, without a handoff between specialist models at every step. Every handoff you remove from an agent pipeline is a failure mode you no longer debug at 2 a.m.

The half-year context also explains the pricing. MiniMax entered a market where standard-tier buyers already had strong choices, so M3 had to undercut on price while overdelivering on modality — and that is precisely the shape of the deal you get. Competitive pressure is a wonderful thing when you are the customer.

How Does MiniMax M3 Fit a Real Workflow?

The simplest way to slot it in: make M3 your default for anything involving mixed media or long agent runs, and keep a premium model on speed dial for the hardest pure-reasoning tasks. On CoreAI you can do that inside a single conversation — switch models mid-thread, attach files and images, and let the history carry over across iOS, Android, and the web under one subscription.

Before committing, run your own bake-off. Open CoreAI Compare, paste a real task — ideally one with a screenshot attached — and run M3 against Qwen3.7 Plus and whatever you use today. Vendor benchmarks are marketing; your own prompt is data. And if the budget end of the market is your hunting ground, our budget AI models roundup maps where M3's cheaper cousins draw the quality line. Writers using M3 for summarization work can also grab CoreAI's free AI Text Summarizer for quick one-off jobs that do not need a full chat.

Frequently Asked Questions

Does MiniMax M3 really accept video input?

Yes. M3 accepts text, image, and video input with text output, and the multimodal capability is native — the model was trained on mixed modalities from the start rather than having vision attached afterward. Native video understanding is still rare among open-weight models, which is a large part of M3's appeal.

Is MiniMax M3 open source?

The weights are openly available on Hugging Face, published within a week of the announcement. You can self-host it or use a hosted version; on CoreAI it is a standard-tier model available under the regular subscription with no API keys required.

How good is MiniMax M3 at coding?

MiniMax reports 59.0% on SWE-Bench Pro, which puts it in serious agentic-coding territory for its price class. Its edge is combining code understanding with visual context — screenshots and screen recordings — plus a 512,000-token output cap for very large single-shot generations. Treat the benchmark as vendor-reported and verify on your own tasks.

What context window does MiniMax M3 support?

As served on CoreAI, the context window is 524,288 tokens with up to 512,000 output tokens per response. MiniMax's own materials cite a 1M-token window for the model; either way, the served window comfortably holds hundreds of pages of mixed input.

Is MiniMax M3 cheaper than Claude or GPT models?

Substantially. At $0.30 per million input tokens and $1.20 per million output, M3 costs a fraction of premium options like Claude Sonnet 5 ($2.00 in, $10.00 out) while adding video input those models lack. Whether the quality gap matters for your tasks is exactly what a side-by-side comparison will tell you.