MiniMax M3: Is the June Open-Weight Star Still Worth Using?

MiniMax M3 launched in June as the leading open-weight model. Five months and several rivals later, here is where it still makes sense.

Disclosure: this guide draws on MiniMax’s announcement, Artificial Analysis evaluations and provider documentation as reported by third parties. We have not self-hosted MiniMax M3. Benchmark figures are vendor-reported unless attributed to an independent tester.

MiniMax released MiniMax M3 on June 1, 2026, and for a while it was the strongest open-weight model available. Artificial Analysis scored it 55 on its Intelligence Index, just ahead of Kimi K2.6 and MiMo-V2.5-Pro, and said it would lead the open field once the weights landed. They did, shortly afterwards.

Five months later the honest answer is more nuanced. MiniMax M3 is no longer the leader; DeepSeek, Meta, Moonshot and Mistral have all shipped newer open models since. But it still offers a combination that is rare at its price: open weights, a million-token context, native image and video input, and API pricing of $0.30 per million input tokens. For long-context multimodal work on a budget, it remains a sensible choice.

Table of Contents

MiniMax M3 network cables plugged into a switch
Long context is cheap to advertise and expensive to serve. MiniMax M3 made it affordable by deciding what not to attend to.

What MiniMax M3 Is

MiniMax is a Shanghai-based lab whose previous model, M2.7, was a well-regarded open-weight agentic model with a 200,000-token context. MiniMax M3 was pitched as the first open-weights system to combine three things that usually come separately: frontier-level coding, a one-million-token context window and native multimodality.

Artificial Analysis lists it at 428 billion total parameters with 23 billion active, a mixture-of-experts design. It accepts text, images and video and returns text. The model ID is MiniMax-M3, and it is case-sensitive on MiniMax’s own API.

It is available through the MiniMax API, MiniMax Code, the company’s agent product, and monthly token plan subscriptions, as well as through OpenRouter as minimax/minimax-m3. The API is compatible with OpenAI’s Chat Completions format, so existing clients work with a base URL change.

MiniMax Sparse Attention Explained

The technical idea behind the long context is what MiniMax calls MiniMax Sparse Attention, or MSA. Standard attention compares every token with every other token, so cost grows with the square of context length. At a million tokens that becomes prohibitively expensive.

Sparse attention cuts the work by having each token attend to a selected subset of others rather than all of them. The art is in choosing the subset so that the tokens that matter are almost always included. Fireworks described MSA as the breakthrough enabling MiniMax M3’s extended context window.

The practical consequence is that long context becomes affordable to serve, which is why MiniMax can offer a million tokens at prices well below Western proprietary models. The trade-off, as with any sparse scheme, is that recall on very long documents depends on the selection working well for your content. Test it with your own long inputs.

One caveat at launch: Fireworks noted that the open-weight release supported up to 500,000 tokens of context, while the hosted API offered the full million. Self-hosters should check which limit their inference engine supports.

MiniMax M3 Pricing

Context band Input per 1M Output per 1M
Up to 512K tokens $0.30 $1.20
512K to 1M tokens $0.60 $2.40

Artificial Analysis calculated a blended rate of $0.22 per million tokens using its standard mix of cache hits, input and output. MiniMax also offered 50% off for the first seven days after launch.

The pricing story on third-party providers moved too. Fireworks initially priced MiniMax M3 at about twice M2.7 on its serverless tier, then permanently dropped it to parity with M2.7 once the open weights were released. Competition between hosts tends to push open-model prices down after weights go public, which is one of the quiet benefits of open releases.

DataNorth calculated that MiniMax M3 was more than fifteen times cheaper on input than leading Western proprietary models at launch. Since then, Western prices have fallen sharply too, so the gap is smaller, but the model remains inexpensive.

MiniMax M3 Benchmarks

Benchmark MiniMax M3 Note
SWE-Bench Pro 59.0 Above GPT-5.5 and Gemini 3.1 Pro at launch
Terminal-Bench 2.1 66.0 Leading among open models at launch
MCP Atlas 74.2 Tool use
GPQA Diamond 93% Graduate-level science
OSWorld-Verified 70% Computer use
AA Intelligence Index 55 Independent, June 8

These were strong numbers in June. VentureBeat’s headline claimed it eclipsed GPT-5.5 and Gemini 3.1 Pro on key benchmarks for a fraction of the cost. The AI Track added an important qualifier: it trailed Claude Opus 4.8 on several directly comparable agent benchmarks. OfficeChai noted that abstract reasoning and true generalisation were less flattering.

All the vendor figures were early release benchmarks. The independent index score of 55 is the most reliable single number, and it placed MiniMax M3 at the top of the open field at the time.

How MiniMax M3 Holds Up Five Months On

The open-model landscape changed fast after June. DeepSeek released V4.1 Flash in September with MIT weights and lower prices. Meta released Muse Spark Open in October, an 816B model. Mistral released Mistral Large 4 in late October. Moonshot’s Kimi K3 posted 85.0 on Terminal-Bench 2.1 in Meta’s comparison table, well above MiniMax M3’s 66.0.

So MiniMax M3 is no longer the default answer to which open model is best. It is now one of several strong options, distinguished less by peak capability and more by its particular mix of features. The question is whether that mix fits your work.

Where MiniMax M3 still stands out is native video input. Most open competitors accept text and images; fewer handle video natively. For workloads that analyse video content alongside long text, it remains one of the few open choices.

It also benefits from maturity. Five months after release, inference engines, quantised builds and fine-tunes have caught up. Running a model with a settled ecosystem is easier than running last week’s release, and that matters in production.

Self-Hosting MiniMax M3

With 428 billion parameters, MiniMax M3 needs server-class hardware to self-host. Simple arithmetic puts 4-bit weights at roughly 214 GB, plus context cache. Lushbinary’s developer guide noted that self-hosting depends on inference engines such as vLLM or SGLang adding support for MSA, since the sparse attention scheme needs specific kernels.

The small active parameter count, 23 billion per token, keeps generation fast once the model is loaded. That makes it a reasonable candidate for high-throughput serving on a multi-GPU node.

Licence terms deserve a direct check. Some trackers list it as Apache 2.0, but confirm the current terms on the official model card before shipping a commercial product. Licence details for open models change between announcement and release more often than you might expect.

Five Reasons to Still Use MiniMax M3

  1. Native video input. Few open models accept video directly. If your product analyses video, this is a real differentiator.
  2. A million-token context at low cost. Up to 512,000 tokens stays at the base rate, which is generous.
  3. A mature ecosystem. Five months of community tooling, quantisations and provider support make it easier to deploy than the newest releases.
  4. Competitive tool use. An MCP Atlas score of 74.2 at launch is solid for agents that call external tools.
  5. Provider competition. With open weights, several hosts serve it, which keeps prices down and gives you options if one provider has an outage.

Where It Falls Short

Peak capability is the obvious limit. Newer open models score higher on coding and agentic benchmarks, and the best closed models are further ahead still. For the hardest coding tasks, MiniMax M3 is not the strongest open option any more.

Abstract reasoning was a weak spot even at launch, according to OfficeChai. If your workload involves novel problem-solving rather than applying known patterns, test carefully.

The sparse attention scheme that makes long context cheap also makes long-context recall something to verify rather than assume. Place key facts at different positions in a long test document and check that the model finds them.

Finally, the data-policy question applies here as with other Chinese-developed models. Some organisations restrict sending data to Chinese-hosted APIs. Self-hosting the weights or using a Western provider that hosts the open model are the usual workarounds.

Moving From M2.7 or Another Model

For teams already on MiniMax’s previous model, the upgrade is the easiest decision in this guide. The API format is the same, several hosts now price the new model at the old one’s level, and the context window jumps from 200,000 tokens to a million. Swap the model name, rerun your evaluation set and watch costs for a week.

Moving from a closed model takes more care. Prompts written for one vendor often rely on that model’s habits, such as how it formats lists or how strictly it follows a system prompt. Expect to adjust a few instructions, and keep the old model available as a fallback while you compare. Pay particular attention to structured output, since small formatting differences can break downstream parsers silently.

Testing Long-Context Recall Properly

Because the main reason to choose MiniMax M3 is its long context, it deserves a proper long-context test before you rely on it.

Build a test document from your own material, ideally a few hundred thousand tokens of the kind of content you will really process. Plant a handful of specific facts at different depths: near the start, in the middle, near the end. Then ask questions that can only be answered from those planted facts.

Run the test several times with the facts in different positions. Long-context models often recall the beginning and end of a document well and lose detail in the middle. Sparse attention schemes can also behave differently depending on the type of content, so a test on legal text tells you little about performance on source code.

Ask multi-hop questions too, ones that require combining two planted facts from different parts of the document. That is harder than simple retrieval and closer to real analysis work.

If recall is weak at your target length, consider a hybrid approach: use retrieval to pull the most relevant sections into a shorter prompt. Often a well-chosen 50,000-token context beats an unfiltered million-token one, and it costs less too.

What Native Video Input Makes Possible

Video understanding is the capability most teams have not tried yet, mainly because so few affordable models offer it. It is worth thinking through what it enables before dismissing it as a niche.

Support teams receive screen recordings of bugs every day. A model that can watch a two-minute recording, describe what the user did and where the error appeared, and match it against known issues saves an engineer a lot of time. Text descriptions of bugs are often vague; recordings are not.

Training and compliance teams have hours of recorded sessions, procedures and inspections. Searching those by content, rather than by file name, becomes possible when a model can describe what happens in each segment. So does checking whether a recorded procedure followed the written one.

Product and research teams record user interviews and usability tests. Summarising them, pulling out quotes and flagging moments of confusion is slow manual work. Combined with a long context for the transcript, a video-capable model can produce a first pass in minutes.

The cautions are the usual ones. Video is expensive in tokens, so long recordings eat budget quickly. Accuracy on fine visual detail varies, so verify anything that matters. And recordings often contain personal data, which makes the self-hosting option attractive for privacy reasons.

Choosing an Open Model in Late 2026

The open-weight field now has enough strong options that the choice depends more on your constraints than on a single leaderboard. A few questions narrow it quickly.

What licence can you accept? MIT and Apache 2.0 are the simplest. Custom licences may require attribution or restrict hosted services. That alone rules some models in or out.

Where may your data go? If Chinese-hosted APIs are off limits, you need either a US or European model, or self-hosting of whichever open weights you prefer.

What hardware do you have? A single consumer GPU means a model around 30B parameters. A multi-GPU server opens up the large mixture-of-experts models.

What modalities do you need? Text only, text and images, or video too. Fewer models cover the last case.

How long are your inputs? If you regularly process hundreds of thousands of tokens, prioritise models whose long-context pricing and recall hold up.

Answer those five, and the shortlist usually comes down to two or three models. Then test on your own tasks. That last step is the one people skip, and it is the only one that tells you what you will actually get.

Who Should Use What

If you are Start with Why Link
Analysing video with long text MiniMax M3 Native video input in an open model See analysis
Wanting the best open coding model Kimi K3 or Muse Spark Open Higher recent coding scores Compare
Chasing the lowest price DeepSeek V4.1 Flash Cheaper and MIT licensed See specs
Already on MiniMax M2.7 MiniMax M3 via your current host Same price on some providers, bigger context See Fireworks

Related on this site: for more context, see DeepSeek V4.1 Flash, Mistral Large 4 and Open-Weight Models in November 2026.

Frequently Asked Questions About MiniMax M3

When was MiniMax M3 released?
MiniMax released it through its API on June 1, 2026, and published the open weights shortly afterwards.

How much does MiniMax M3 cost?
$0.30 per million input tokens and $1.20 output for contexts up to 512,000 tokens, rising to $0.60 and $2.40 between 512,000 and one million.

How big is the model?
Artificial Analysis lists 428 billion total parameters with 23 billion active per token, in a mixture-of-experts design.

What is MiniMax Sparse Attention?
MiniMax’s sparse attention scheme, which lets each token attend to a selected subset of the context. It is what makes the million-token window affordable to serve.

Is it still the best open model?
No. Newer open models from DeepSeek, Meta, Moonshot and Mistral score higher on several benchmarks. It remains competitive, especially for long-context and video work.

Can I self-host it?
Yes, with server-class hardware and an inference engine that supports MSA. The open release initially supported up to 500,000 tokens of context.

Verdict

Use MiniMax M3 when you need an open model that handles long documents and video together at a low price. Its mature ecosystem and competitive tool use make it a dependable choice for that niche five months after launch.

The caveat is that it is no longer the strongest open model. Newer releases beat it on coding, abstract reasoning was always a weak spot, and the benchmark figures were early vendor numbers. Verify long-context recall before you depend on it.

Next step: build a long test document from your own content, plant facts at several depths, and compare MiniMax M3 against one newer open model on recall, accuracy and cost.

Sources and Further Reading

On this site: serve it with help from LLM inference servers compared, combine it with retrieval-augmented generation, compare smaller models in best local LLMs, and see the field in AI models compared.