A business owner reads that a Chinese AI lab has released a model reporting frontier-level coding benchmark results, at a fraction of what OpenAI or Anthropic charge, and the weights are free to download. The natural next thought is: can we just run this ourselves and cut our AI costs to almost nothing?
The honest answer is more nuanced than either the excited headlines or the dismissive ones suggest. MiniMax M3, released by the Shanghai lab MiniMax on June 1, 2026, is a genuinely capable open-weight model. But "open weight" here does not mean what it sounds like on first read: it is not free for unrestricted commercial use, and self-hosting it is not a project you run on a laptop over a weekend. It also is not, for most SMBs, a reason to avoid the model entirely.
This article covers what M3 actually is, the real difference between using it through an API and self-hosting it, what the license actually requires, and where it makes sense for a European small or mid-sized business.
What MiniMax M3 actually is
M3 is a mixture-of-experts model with roughly 428 billion total parameters, of which about 23 billion are active for any given token. It supports a context window of up to 1 million tokens, built on an architecture MiniMax calls Sparse Attention, and it was trained on text, images, and video from the start rather than having multimodal capability added afterward. That combination, long context plus native multimodality plus strong coding and agent performance in a single open-weight release, is what generated most of the attention around the launch.
MiniMax reports frontier-level coding results for M3, including 59.0% on SWE-Bench Pro, which the company says surpasses GPT-5.5 and Gemini 3.1 Pro on that specific benchmark. Results vary by benchmark, though: on another MiniMax-published evaluation, M3 scored lower than both GPT-5.5 and Claude Opus 4.7. Worth noting: these figures come from MiniMax's own evaluation, and MiniMax's published comparisons used Claude Opus 4.7 as the reference point rather than the more recent Opus 4.8, which shipped days before M3's launch. Independent, third-party benchmark verification for M3 is still limited as of this writing. Treat the headline numbers as a reasonable, benchmark-specific signal of capability, not as a settled overall ranking.
It's also worth keeping in mind that this market moves fast. M3 launched into a landscape that has already shifted again since, newer GPT-5.6 and Claude Opus 5 releases are now available, so any benchmark or price comparison in this article should be read as a point-in-time snapshot rather than a permanent ranking.
| Spec | MiniMax M3 |
|---|---|
| Total parameters | About 428 billion |
| Active parameters per token | About 23 billion |
| Context window | Up to 1 million tokens |
| Multimodal input | Text, image, and video |
| API pricing | Roughly $0.30 to $0.60 per million input tokens, $1.20 to $2.40 per million output tokens, depending on context length |
| Weights available | Yes, on Hugging Face |
| License | MiniMax Community License, not MIT or Apache |
Two ways to run it, and they are not close
There are two paths to using M3, and the gap between them is bigger than it looks from a spec sheet.
The API path is a base URL and an API key away from working in any setup already built for OpenAI-compatible tooling. No infrastructure to manage, elastic capacity, and, on published standard rates as of August 2026, meaningfully lower per-token pricing than several flagship closed models:
| Model | Input price per million tokens | Output price per million tokens |
|---|---|---|
| MiniMax M3, standard context | $0.30 | $1.20 |
| MiniMax M3, long context | $0.60 | $2.40 |
| GPT-5.5 | $5.00 | $30.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
M3 is materially cheaper than these particular flagship models on standard API token pricing, though total task cost also depends on actual token usage, caching, reasoning behaviour, and output quality for the specific workload, not list price alone. It's also worth noting that pricing across the market moves quickly: some newer or lower tiers from other providers are priced closer to M3 than these flagship rates suggest, so this table is a point-in-time comparison, not a permanent ranking. We go into how these token costs translate into a real project budget in How Much Does AI Automation Cost in 2026?. For most businesses evaluating whether M3 is useful at all, the API is the sensible starting point.
The self-hosting path is real, but it is not a local install. A 428-billion-parameter model needs substantial accelerator memory and production-grade inference infrastructure to serve at usable speed, generally multiple data-center-class GPUs working together. The exact configuration depends on precision, quantization, target context length, the serving framework in use, and your throughput requirements, so treat any specific GPU count or storage figure you see quoted online as one possible setup rather than a fixed requirement. This is server-room or cloud-GPU-rental territory, not a business laptop or a single workstation. A handful of enthusiasts have reported running heavily quantized versions on very high-memory Mac Studio configurations, but at reduced quality and well below the model's full 1-million-token context capacity.
For a European SMB without an existing GPU infrastructure team, the realistic self-hosting options are renting dedicated cloud GPU capacity or working with a partner who already operates that infrastructure. Buying and running the hardware in-house rarely makes financial sense below a fairly high, sustained volume of usage.
What this means if you're evaluating open-weight models for your business
For most SMBs evaluating open-weight infrastructure, the goal is not to become an AI infrastructure operator. The real question is whether an open-weight model changes the cost structure or gives access to something the hosted API version of Claude or GPT doesn't. For the large majority of SMB automation, coding assistants, customer support workflows, internal document tools, the hosted API is the practical starting point regardless of which model family you choose, and self-hosting only becomes worth the operational overhead once volume, data control requirements, or contractual constraints justify it.
Where self-hosting M3 specifically earns its complexity is when a business has a genuine reason to keep model inference inside its own infrastructure rather than sending requests to any third party's servers, including MiniMax's own API. We look at that question with a short filter before recommending either path.
A framework for deciding whether to self-host
This is a Kubera AI planning heuristic, not a universal benchmark, meant to shortcut the decision rather than replace a proper infrastructure assessment.
The Kubera Open-Weight Deployment Filter asks four questions:
Does your use case require data to stay inside infrastructure you control? If yes, self-hosting is worth evaluating regardless of which open-weight model you pick. If no, the hosted API is almost always simpler and cheaper to start with. Is your usage volume high and sustained, not occasional? Self-hosted infrastructure has a fixed cost whether you use it or not. Below a certain steady volume, API pricing usually wins on cost even before counting engineering time. Do you already have, or are you willing to build, the technical capacity to operate multi-GPU infrastructure? This includes monitoring, security patching, and capacity planning, not just the initial setup. Does your business fall under the license's commercial threshold, and are you prepared to meet its conditions? This is worth checking before either path, but it matters most if you plan to build a product around the model rather than use it internally.
If the honest answers land mostly on "no," the API is the right starting point, and it can always be revisited later. If you're weighing this alongside other first-project decisions, the same kind of scoring approach we use in How to Choose Your First AI Automation Project applies here too.
The licensing detail almost everyone glosses over
"Open weight" and "open source" are often used interchangeably in coverage of these releases, and the difference matters here. M3 ships under the MiniMax Community License, not MIT or Apache 2.0. Non-commercial use is permitted with the copyright notice retained. For commercial use, the license requires displaying a "Built with MiniMax M3" attribution, and it draws a line at $20 million in annual revenue from the relevant product or service: businesses below that threshold need to send MiniMax a one-time notice, while businesses above it need separate prior written authorization before commercial deployment. Self-hosting the weights does not remove these conditions.
None of this is unusual for the current generation of Chinese open-weight releases, MiniMax's earlier M2.7 model carried similar restrictions, but it is a meaningfully different arrangement from a permissive license, and it is easy to miss if you only read that the weights are "open" and stop there. If you are planning to build a commercial product on top of M3 rather than use it as an internal tool, reviewing the current license text before committing engineering time is worth the hour it takes.
Why self-hosting can matter specifically for a European business
Beyond cost and volume, there is a data-handling angle worth being precise about, rather than assuming. MiniMax's own API privacy documentation describes cross-border data storage on US-based cloud infrastructure and references the EU-US Privacy Framework, not storage inside China. That is a meaningfully different picture from the assumption some coverage of Chinese AI labs leads with, and it's worth reading the current privacy documentation directly rather than relying on general assumptions about where a Chinese company's servers sit.
That said, MiniMax being a China-headquartered company is a separate question from where a given API request is physically processed, and it's a distinction worth keeping straight rather than collapsing into a single claim either way. Before sending personal or regulated data through any third-party API, including MiniMax's, the sensible step is to review its current documentation on storage location, international transfers, data retention, processor terms, and contractual safeguards, since these details can change and matter more than the vendor's country of origin on their own.
Self-hosting the open weights on your own servers, whether on-premises or on a European cloud provider, does give you more direct control over inference and data flow: the model runs on infrastructure you control, and request data doesn't need to leave it at all. Whether that additional control is worth the operational overhead depends on your specific data-handling requirements, not on an assumption about the API's location. We cover the broader data-handling side of this question in Is Your Business Data Safe with AI?.
Pros and cons for a European SMB
| Consideration | What it looks like in practice |
|---|---|
| Cost, via API | Cheaper per token than some current flagship closed models on standard pricing, useful for high-volume, less judgment-heavy tasks, though the gap to cheaper tiers from other providers is narrower |
| Cost, self-hosted | Only competitive at sustained high volume, once GPU rental or ownership costs are factored in |
| Capability | Strong coding and agentic performance on several vendor-reported benchmarks, mixed results on others, with independent verification still limited |
| Context and multimodality | Genuinely useful for large codebases, long documents, and mixed text-image-video tasks |
| Licensing | Not a permissive open-source license, commercial use has explicit conditions and a revenue threshold |
| Data control | Self-hosting gives direct control over inference and data flow; the hosted API's actual data handling should be checked against MiniMax's current privacy documentation rather than assumed |
| Ecosystem maturity | Younger than Claude or GPT tooling, fewer established integrations and support channels |
Where MiniMax fits against the alternatives
M3 is one of several Chinese labs shipping strong open-weight models this year, alongside DeepSeek and Moonshot AI's Kimi family, which we cover in more depth separately. Compared with closed, hosted-only models like Claude and GPT, which we've compared in Claude vs ChatGPT vs Gemini vs Qwen vs DeepSeek, the trade-off with M3 is fairly consistent: lower cost and the option to self-host, against a less mature ecosystem, licensing conditions to track, and benchmark claims that still need independent confirmation.
What we would actually choose
For a business that wants to experiment with a capable, low-cost model for coding assistance or high-volume content tasks, and does not have a specific reason to avoid third-party infrastructure, the M3 hosted API is a reasonable and inexpensive option to test alongside your existing tools. For a business with a genuine data-residency requirement, regulated data, contractual restrictions on where processing happens, or a client base that specifically expects EU-based data handling, self-hosting an appropriate open-weight model on European infrastructure becomes worth evaluating, provided the volume justifies the fixed cost. M3 is one reasonable candidate for that path, not automatically the best one; the right choice depends on how its license terms, benchmark profile, and ecosystem maturity compare with other open-weight options for your specific workload. For most day-to-day automation with moderate volume and no specific data-sovereignty constraint, we'd still default to the established hosted providers first and treat M3 as a cost-optimization layer for specific, well-defined tasks rather than a wholesale replacement.
FAQ
Is MiniMax M3 free to use?
The API is paid, at low per-token rates. The weights can be downloaded and used commercially, but only under the terms of the MiniMax Community License, which includes attribution and notice requirements.
Can I run MiniMax M3 on a regular office server?
No. Serving the full model at practical production throughput generally requires substantial accelerator memory and multi-GPU infrastructure. A typical office server without dedicated AI accelerators cannot run it at usable speed.
Is MiniMax M3 open source?
Not in the conventional sense. The weights are downloadable and inspectable, but the license is a custom Community License with commercial conditions, not a permissive license like MIT or Apache 2.0.
How does MiniMax M3's cost compare to Claude or GPT?
On published standard rates as of August 2026, M3 costs $0.30 per million input tokens and $1.20 per million output tokens, well below GPT-5.5's $5.00 and $30.00 or Claude Opus 4.8's $5.00 and $25.00. Newer, cheaper tiers from other providers have narrowed that gap since, so treat this as a point-in-time comparison and check current pricing before deciding.
Should my business self-host MiniMax M3?
Usually only if you have sustained high usage volume, existing or planned multi-GPU infrastructure, and a specific reason, such as data residency, to keep inference inside infrastructure you control. Otherwise, the hosted API is simpler and cheaper to start with.
Are MiniMax M3's benchmark scores reliable?
The published scores are vendor-reported and use MiniMax's own evaluation setup, including a comparison baseline that predates the most recent competing model releases. Results also vary by benchmark, M3 leads on some and trails on others. They're a useful signal, not an independently verified ranking.
Does MiniMax M3 support images and video, or just text?
Both. It was trained on mixed text, image, and video data from the start, rather than having multimodal support added to a text-only base model.
What's the difference between MiniMax M3 and MiniMax M2.7?
M2.7 is MiniMax's previous flagship, text-only, with a smaller context window. M3 adds native multimodality and extends the context window significantly, alongside architecture changes aimed at making long-context inference more efficient.
Is my data safer using the MiniMax API or self-hosting the weights?
Self-hosting keeps your request data inside infrastructure you control, which is a real advantage if data residency specifically matters to your business. MiniMax's own API privacy documentation describes cross-border storage on US-based cloud infrastructure rather than storage in China, so this isn't a simple "API equals less safe" comparison. It's worth reviewing MiniMax's current privacy terms directly, on storage location, transfers, and retention, before sending personal or regulated data through the API.
Does the $20 million revenue threshold in the license apply to a small agency?
For most SMBs, no, but the license still requires a one-time notice to MiniMax for commercial use below that threshold, plus attribution. It's worth reading the current license text directly before deploying commercially, since terms can be updated.
If you're weighing whether an open-weight model like MiniMax M3 fits your automation stack, or whether the hosted API is the better starting point for your situation, that's exactly the kind of decision worth mapping out before committing engineering time to either path.