What Is MiniMax M3? The First Open Model to Do All Three at Once
Before MiniMax M3, getting frontier-level coding, a million-token context window, and real multimodal understanding out of one AI system meant paying for a closed model — Claude Opus, GPT-5.5, or Gemini. MiniMax M3, released June 1, 2026, was the first open-weight model to combine all three. The developer community has been arguing about it ever since — not because it's hype, but because parts of it hold up. Here's what's real, and what's still pending independent confirmation.
MiniMax M3 combines frontier coding, a 1M-token context window, and native multimodal input in one open-weight release.
Coding, Long Context, and Vision in One Open Model
MiniMax M3 is an open-weight AI model from MiniMax, launched June 1, 2026. It's a 428-billion-parameter mixture-of-experts model with about 23 billion active parameters, built on MiniMax's own Sparse Attention (MSA) architecture that makes a 1-million-token context window economically viable. It's MiniMax's first M-series model with native multimodal support, adding image and video input on top of the text-only capabilities of its predecessor, MiniMax M2.7.
MiniMax's pitch is specific: this is the first open-weight model to bring frontier-level coding, million-token context, and native multimodality together in a single system. On the Artificial Analysis Intelligence Index, M3 scores 55 — just ahead of open-weight peers like Kimi K2.6 (54) — and the firm noted that once weights were released, it would be the leading open-weight model by that measure.
Real Demos, Honestly Mixed Independent Testing
12-hour autonomous run
M3 reproduced an ICLR 2025 Outstanding Paper independently, running nearly 12 hours and producing 18 commits and 23 experimental figures without human intervention.
CUDA kernel optimization
In a separate demo, M3 optimized a CUDA kernel from 7.6% to 71.3% hardware utilization over 24 hours with zero human intervention.
Independent Intelligence Index score
Artificial Analysis independently scored M3 at 55 on its Intelligence Index, ahead of Kimi K2.6 and MiMo-V2.5-Pro, though behind Claude Opus 4.8 and GPT-5.5 on GDPval-AA.
Mixed community coding results
Early independent YouTube evaluations found a genuine leap over M2.7, but inconsistent wins against Claude or GPT-5.5 in actual coding sessions — some users reported DeepSWE pass@1 well below MiniMax's own headline figures.
That gap matters: the marquee demos were run on MiniMax's own infrastructure with its own agent scaffolding, so treat them as a strong signal rather than a fully independent verdict. The Artificial Analysis Intelligence Index score, by contrast, is a genuinely third-party number — and it's a solid one for an open-weight model of this size.
What Actually Changed
MiniMax reports meaningful gains across most evaluations from M2.7 to M3, with one small regression on SciCode.
| MiniMax M3 | MiniMax M2.7 | |
|---|---|---|
| Intelligence Index | 55 | 50 |
| Multimodal input | Yes — image, video | No — text only |
| Context window | Up to 1M tokens | Smaller, text-focused |
| HLE score | 37% | 28% |
| GPQA Diamond | 93% | 87% |
MiniMax reports M3 improving on M2.7 across most evaluations — a 9-point jump on HLE, 6 points on GPQA Diamond, 5 on AA-LCR, and 7 on IFBench — with a small regression on SciCode (47% down to 45%). One notable caveat from Artificial Analysis: M3 attempts only about 31% of questions on the AA-Omniscience benchmark, the lowest attempt rate among current peers, which produces both a low hallucination rate and a low accuracy score — the model appears tuned to abstain rather than guess.
Why Sparse Attention Is What Makes 1M Context Affordable
A million-token context window sounds like a spec sheet number, but it's actually the hardest part of this release to pull off economically. Standard "dense" attention — the mechanism most transformer models use to let every token look back at every previous token — scales quadratically with context length: double the context, and the compute cost roughly quadruples. That math is exactly why most closed frontier models cap their practical context well below what they technically support, and why running a long-context request has historically cost noticeably more per token than a short one.
MiniMax's Sparse Attention (MSA) architecture is the reason M3 can offer a 1M-token window without that cost curve becoming unworkable. Instead of every token attending to every other token, MSA selectively restricts which token pairs actually compute attention against each other, cutting the effective compute load for long sequences while aiming to preserve most of the accuracy dense attention would have delivered. It's not a new idea in AI research generally — sparse and linear attention variants have circulated in papers for years — but shipping one at 428B parameters, in a model that also has to hold up on demanding coding and reasoning benchmarks, is a harder engineering bet than publishing a research paper on the technique.
The practical upshot for anyone actually using M3: the 512K-to-1M context tier costs exactly double the base tier per the pricing table above, not the much steeper multiple a dense-attention model would need to charge for the same jump. If your workload genuinely needs to hold a large codebase or a long document in context at once — rather than context length being a "nice to have" you rarely touch — that pricing structure is the concrete reason M3 is worth evaluating over a dense-attention alternative with a similar headline window size — the difference compounds fast once you're running agent sessions that accumulate context over many turns rather than a single one-shot prompt, where every added turn would otherwise multiply the cost penalty of dense attention.
Who MiniMax M3 Is For
- Genuinely rare combination: frontier coding, 1M context, and multimodality, all open-weight
- Independently confirmed Intelligence Index score ahead of several open-weight peers
- Priced well below closed frontier models — $0.30/$1.20 per million tokens at base tier
- Strong demonstrated long-horizon autonomy on real, verifiable tasks
- Headline demos ran on MiniMax's own infrastructure and scaffolding
- Community coding tests show inconsistent wins against Claude and GPT-5.5
- Low attempt rate on some benchmarks trades accuracy for a lower hallucination rate
- As a China-based lab, data-residency questions apply for regulated industries
Frequently Asked Questions
What is MiniMax M3?
An open-weight, 428B-parameter mixture-of-experts model from MiniMax, launched June 1, 2026, with a 1M-token context window and native multimodal support.
Is it really the first to combine all three capabilities?
MiniMax's specific claim — frontier coding, 1M context, and multimodality in one open-weight model — was largely confirmed by independent evaluators, though some coding claims were still pending broader verification.
How much does it cost?
$0.30/$1.20 per million input/output tokens up to 512K context, rising to $0.60/$2.40 for 512K–1M context, via MiniMax's API. Weights are also downloadable for self-hosting.
More AI Model Guides, Written Clearly
Browse our full library of AI model explainers, updated as new releases land.
More from the creator
Explore two other projects from the same creator.