At a glance
- Maker: MiniMax
- Released: 2026
- Parameters: 427B
- Context window: 1M
- Size: Very large
- Memory: Handles very long documents
- Good at: Long agent tasks and video
- Supports: tools, images, zero data retention
- License: Apache 2.0
- Upstream API price: $1.20 per million tokens
Announced in MiniMax M3: Frontier Coding, 1M Context, Native Multimodality.
MiniMax M3 is paused while we bring up more capacity. The picker on the chat page shows what is ready right now, and the model returns once capacity is up.
What is MiniMax M3?
MiniMax M3 is the multimodal flagship of MiniMax, the Shanghai lab best known outside China for its Hailuo video models — and M3 wears that heritage openly. Its announcement is titled Frontier Coding, 1M Context, Native Multimodality, which reads like a checklist of the current flagship fashion, and the model delivers on all three: it is built for coding and long agent work, it holds a million tokens, and pictures and video go in as native input.
The scale is serious without being extravagant: 427 billion parameters, a one-million-token context window — roughly 750,000 words — tool support, zero-data-retention terms, and a $1.20 per-million-token price that undercuts most of the flagship field. Its stated specialities are long agent tasks and video, the two places its multimodality and context combine most usefully.
One honest note on availability: M3 is currently paused while UsingOpen brings up more capacity, so the picker may show it greyed out. The page and the model remain; capacity returns. That is the ordinary rhythm of free access to popular open models, and it is worth checking back rather than writing the model off.
What can you do with MiniMax M3?
When M3 is selectable, it is a strong pick for:
- Long agent tasks: many-turn tool workflows that stay coherent over hours.
- Coding across big projects, with the million-token window holding everything.
- Questions about pictures and video — native multimodality, not a bolt-on.
- Summarising marathon inputs: whole document sets and long transcripts at once.
- Agent-plus-video work: pipelines that watch things and act on them.
- Flagship-quality drafting at less than flagship prices.
How does MiniMax M3 compare?
At $1.20, M3's natural rival is Thinking Machines' Inkling Small — same price, also multimodal with tools, but a smaller context (524k) and a quick-answer temperament versus M3's long-agent brief. Against Z.ai's GLM 5.3 Flash ($0.25), M3 is the escalation: five times the price for more scale and staying power on very long jobs.
Against the true heavies — Kimi K3 at $8.50, Qwen3.8 2.4T at $6.00 — it concedes some peak depth in exchange for a much gentler price and native video. For long-horizon multimodal agents, that trade frequently favours M3.
Is MiniMax M3 right for you?
M3 is the pick for long-horizon multimodal agents: tool loops that run for hours, coding sessions with video or picture material mixed in, jobs where a million tokens of context and native sight both pull weight. At $1.20 per million tokens it undercuts most of the flagship field while behaving like part of it, and the Hailuo video heritage shows in how naturally it handles material that other models treat as an afterthought.
Its current pause complicates the recommendation — the picker shows it as adding capacity — but that is the ordinary rhythm of popular free models, and it returns. The very demand that exhausts the capacity pool is a decent signal about the model. If your timing is tight, Inkling Small matches the price with a different temperament, and the MiMo models cover the multimodal middle at lower cost.
What M3 is genuinely good at is staying coherent: long agent runs are its stated speciality, and the combination of native video understanding with tool support makes it one of the few models here that can watch something and act on it. Teams building monitoring-style or media-style pipelines should keep it on the shortlist, pause or no pause — and check the picker, because the model people keep exhausting capacity for is usually worth the wait.
Frequently asked questions
Why is MiniMax M3 paused on UsingOpen?
The model is popular enough that the free capacity pool ran dry — the picker shows it as adding capacity. It will return; that is the ordinary rhythm of free access to open models. The chat page always shows live status. The pattern is well established across this catalogue — GLM 5.2 is paused for the same reason, and both return. Popular free models and capacity are a repeating story, not a tragedy.
What is the context window of MiniMax M3?
One million tokens — roughly 750,000 words. Long agent runs, whole codebases and marathon transcripts fit in a single session. At a million tokens it shares the top tier of context with the heavyweights, at a fraction of their price.
Does MiniMax M3 support image input?
Yes — pictures and video, natively rather than through a separate vision stage. Video understanding is one of its two stated specialities, a heritage of the Hailuo video line. Native multimodality means the picture or video is part of the model’s own context rather than a pre-processed description of one, which shows in how precisely it can quote what it saw.
Is MiniMax M3 good for agent work?
Yes — long agent tasks are its stated speciality, alongside video. Tool support and the million-token window are both aimed at workflows that run for hours rather than minutes. The million-token window is what makes long runs survivable: the model can keep the original brief in view even hundreds of turns later.
Who makes MiniMax M3?
MiniMax, the Shanghai lab behind the Hailuo video models. Its announcement for M3 is titled Frontier Coding, 1M Context, Native Multimodality — a fair summary of the build. The Hailuo line gave MiniMax unusual depth in video, and M3’s comfort with video input reads like a company playing to its strengths.
Is MiniMax M3 free to chat with on UsingOpen?
Yes, whenever capacity allows. No account, no payment, history saved only on your device. When the model is paused, the picker shows it greyed out until capacity returns. Set the page aside and check back — capacity pools refill, and the model is worth the wait when it does.
Related models on UsingOpen
If MiniMax M3 is not quite the right shape, these neighbours on UsingOpen are worth a look:
- GLM 5.3 Flash by Z.ai: the fast default — 18B of 320B parameters awake per token, a million-token window and picture input.
- Inkling Small by Thinking Machines: the smaller Thinking Machines model — 12B of 276B awake for quick multimodal answers.
- Kimi K3 by Moonshot AI: the biggest model on UsingOpen — 2.8T parameters, built for hard coding.
- MiMo V2.6 Pro by Xiaomi: the trillion-scale Xiaomi flagship for demanding multimodal work.
Head-to-head comparisons
See how MiniMax M3 lines up with the models people most often weigh it against:
- MiniMax M3 vs Qwen3.8 Flash — MiniMax's video-and-agent model against Qwen's faster, cheaper all-rounder.
MiniMax M3 will show in the picker as adding capacity until the pool is topped up — other models remain selectable in the meantime.
Chat with MiniMax M3 free on UsingOpen — no account, no sign-up
← All open-weight models · Compare models · Free AI chat · FAQ