A developer on your team pastes an entire codebase into a context window. Not a summary of it, not a RAG-retrieved slice - the whole thing. Until June 1, 2026, doing that with an open-weight model meant paying a compute bill that made the exercise academic. MiniMax M3 changes the math.
MiniMax M3 is an open-weight mixture-of-experts model from the Shanghai lab MiniMax, released June 1, 2026. It has 428 billion total parameters with roughly 23 billion active per token, a 1-million-token context window, and native image and video input.
MiniMax calls it the first and only open-weight model to bring all three together. That claim is worth testing - and so is the architecture that makes it possible.
What MiniMax Sparse Attention actually does
Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. MSA is MiniMax's answer to that constraint, and it is more specific than the usual "we use sparse attention" handwave.
MiniMax Sparse Attention is a blockwise sparse attention built upon Grouped Query Attention. A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution. The kernel-level implementation matters: KV blocks serve as the outer loop, and queries that map to each block are aggregated per pass, so each block is read from memory only once and memory access remains contiguous.
MiniMax reports this design runs more than 4× faster than open-source alternatives including Flash-Sparse-Attention and flash-moba.
The result in numbers: 1/20th the per-token compute of M2 at 1M-token context, with 9x faster prefill and 15x faster decoding, all without compressing key-values or sacrificing precision.
The practical translation of that ratio is the figure most coverage skips. 1M context inference on M3 costs roughly what 50K context inference costs on a comparable dense model. That is the number that makes whole-codebase analysis viable: you are not paying a 20x compute premium to read an entire repo.
What the benchmarks say - and where to be skeptical
MiniMax M3 is the first open-weight model to combine frontier-level coding performance, a 1-million-token context window, and native multimodal capabilities in a single architecture. The model scores 59.0% on SWE-Bench Pro, surpassing both OpenAI GPT-5.5 and Google Gemini 3.1 Pro on this widely used software engineering benchmark.
Independent benchmarking firm Artificial Analysis put M3 at 55 on its Intelligence Index - MiniMax-M3 scores 55 on the Artificial Analysis Intelligence Index. Once the weights are released, it will be the leading open weights model. M3 is MiniMax's first multimodal M-series model, adding image and video input and a 1M token context window over the text-only MiniMax-M2.7.
But read the eval methodology before you route production traffic. Benchmark comparisons surpassing GPT-5.5 and Gemini 3.1 Pro come from MiniMax's own evaluations and warrant independent third-party validation. The standout demonstrations - a 12-hour paper reproduction, a 24-hour CUDA optimization - are described as MiniMax's internal evaluations; reproducibility by external teams has not yet been independently confirmed.
There is also a specific gap on raw text intelligence. On raw text intelligence, M3 is a mid-pack open-weight model; its separation comes from multimodality and price, not benchmark rank. If your workload is pure text reasoning, GLM-5.2 or DeepSeek V4 Flash are still worth comparing directly.
| Benchmark | M3 (vendor) | Independent note |
|---|---|---|
| SWE-Bench Pro | 59.0% | Vendor scaffolding; independent pending |
| SWE-Bench Verified | 80.5% | Per Morph API docs |
| GPQA Diamond | 93% | Up from M2.7's 87% |
| AA Intelligence Index | 55 | Independent; #1 open-weight at weights release |
| GDPval-AA | ~1670 | Level with Claude Sonnet 4.6 (max) |
The open-weight vs. open-source distinction that matters
M3 is open weight - meaning the model parameters are downloadable from Hugging Face - but it is not fully open source. Training data, the data pipeline, and RL recipe are not published. Weights went live on Hugging Face by June 7, and the technical report landed on arXiv on June 11.
For most teams, this is fine - you get self-hosting, auditability of the weights, and freedom from API dependency. For teams with compliance requirements around training data provenance, it is a constraint worth noting before you commit infrastructure.
Input pricing starts at around $0.30 per million tokens, with a blended cost as low as $0.06 per million with cache optimization. At that price point, running a 1M-token context job - reading a 750,000-token codebase - costs under $1 in input tokens via the API. That changes the feasibility calculation for whole-repo analysis workflows.
Beagle in action#eng-platform, 11:02amThe 12-hour run: a demo worth examining closely
The most striking capability demonstration MiniMax ran was not a benchmark. M3 was tested autonomously reproducing an ICLR 2025 Outstanding Paper, running for nearly 12 hours, producing 18 commits and 23 experimental figures without human intervention. A second demo had the model optimizing CUDA kernels unattended.
These are MiniMax's own evaluations, and external teams have not yet replicated them. What they do illustrate is the architectural requirement that MSA satisfies: a multi-hour agentic session accumulates tool call outputs, intermediate results, and working state across the context window. Without a 1M-token window - and without the compute cost being manageable - you cannot run that kind of session without aggressive context trimming that loses information.
The 1M window is not a vanity number here. It is what makes long-horizon agentic work coherent rather than amnesiac.
MiniMax M3 open-weight model: common questions
What is MiniMax M3 and when was it released?
MiniMax M3 is an open-weight 428B mixture-of-experts model from Shanghai-based MiniMax, released June 1, 2026, with weights published on Hugging Face by June 7. It combines a 1M-token context window, frontier-level coding benchmarks, and native image and video understanding in a single downloadable model.
How does MiniMax Sparse Attention make 1M context affordable?
MSA replaces standard quadratic attention with a blockwise sparse mechanism. At 1M tokens, per-token attention compute drops to 1/20th of full attention, with 9× faster prefill and 15× faster decoding. The practical result: running M3 at 1M context costs roughly what running a comparable dense model at 50K context costs.
How does M3 compare to GLM-5.2 or Kimi K2.5 for coding?
For agentic coding specifically, M3 scores 59% on SWE-Bench Pro (vendor-run). GLM-5.2 leads on several independent long-horizon coding benchmarks, while Kimi K2.5 hits 76.8% on SWE-Bench Verified. M3's differentiation is the combination of 1M context, multimodality, and low price - not raw benchmark rank on text-only tasks.
Is MiniMax M3 actually open source?
No. M3 is open-weight: the model parameters are downloadable under MiniMax's license from Hugging Face, and the technical report is public on arXiv. Training data, the RL recipe, and the data pipeline are not released. This matters for compliance teams that need training data provenance, but not for most self-hosting use cases.
What hardware does M3 require to self-host?
M3's 428B total parameters mean the full weight set must fit in VRAM for interactive inference - you cannot page unselected experts from CPU without latency penalties. A two-node H100 or H200 setup is the practical minimum for full-precision serving. MXFP8 quantization via SGLang reduces the footprint; the NVIDIA technical blog has a verified deployment config for 8-GPU nodes.