MiniMax shipped M3 on June 1, 2026, billing it as the first open-weight model to combine frontier coding, a one-million-token context window, and native multimodality in a single architecture - and posted a 59% score on SWE-Bench Pro at $0.60 per million input tokens. That combination of claims is worth sitting with for a moment, because none of those three things is routine in the open-weight world, and at least one of the numbers deserves real scrutiny before you route production traffic through it.
The short version: M3 is a meaningful release. It is also not the clean headline its launch blog suggests.
What MiniMax M3 actually is
M3 is a 428B total / 23B active mixture-of-experts model with a 1M-token context window and native multimodal support, available via the MiniMax API at roughly $0.30 input / $1.20 output per million tokens at promotional pricing.
The architectural story here is MiniMax Sparse Attention (MSA), which the company says delivers 15.6× faster decoding and 9.7× faster prefill compared to M2 at million-token contexts. That matters practically: a 1M-token window is only useful if latency at that context length is actually tolerable. Most prior attempts at that context scale have been technically possible but economically painful to operate. M3's 1M-token context is five times larger than its predecessor M2.7.
MiniMax demonstrated three long-horizon tasks at launch: an autonomous reproduction of an ICLR 2025 research paper in 12 hours (producing 18 commits and 23 figures), a 24-hour CUDA kernel optimization run that improved FP8 hardware utilization from 7.6% to 71.3%, and a model-training task where M3 scored 0.37 on PostTrainBench by training another model end to end. These are the kind of demonstrations that make an impression. They are also exactly the kind of cherry-picked showcase that every major model launch runs - chosen because they worked, not because they represent median performance.
The benchmark numbers: what holds and what doesn't
M3 scores 59.0% on SWE-Bench Pro, surpassing both GPT-5.5 and Gemini 3.1 Pro on this software engineering benchmark, according to MiniMax's published figures.
The number to hold in your head: 59% on SWE-Bench Pro puts M3 ahead of GPT-5.5 on that specific eval - but the gap to the true closed-source frontier is real. Claude Opus 4.7 scores 87.6% on SWE-Bench Verified, and GPT-5.5 leads on Terminal-Bench. A single benchmark position tells you almost nothing about which model to pick for your actual task.
The caveat that every honest review of this launch flags: the benchmark scores are company-reported and run on MiniMax's own infrastructure, and the open weights had not been released at launch.
MiniMax's own technical report notes that SWE-Bench Verified was tested on internal infrastructure using Claude Code as the scaffolding. Running SWE-Bench with Claude Code as the scaffolding agent and calling the result an M3 score is a common industry practice - but it means you are measuring a system, not a model in isolation. The number is real; what it isolates is not.
Open weights went live on Hugging Face by June 7, 2026 , which resolves the initial concern about weights being promised but absent. Independent community benchmarks are now running.
On the price side, M3 beats GPT-5.5 on SWE-Bench Pro while costing roughly 12× less on input and 12.5× less on output. That differential is significant enough to change the economics of any pipeline running high token volumes, assuming the quality holds on your specific workload.
Where M3 fits against the open-weight field
The Artificial Analysis Intelligence Index (v4.1) places GLM 5.2 as the top open-weight model at 51, ahead of NVIDIA Nemotron 3 Ultra (48), MiniMax M3 (44), DeepSeek V4 Pro (44), and Kimi K2.6 (43) - with the leading closed models about five points above that. M3's headline SWE-Bench Pro position and its overall intelligence index position tell two different stories. Pick the benchmark that matches your workload type.
DeepSeek V4, Kimi K2.6, Qwen 3.6 Plus, GLM 5.1, and several others have all closed the gap on closed-source frontier models in ways that matter for the actual work: multi-step task completion, tool call accuracy, and recoverable failure modes. M3 sits in a competitive pack, not alone at the top.
M3 trades roughly ten SWE-Bench Pro points for an output token that is around twenty times cheaper than Opus 4.8, plus open weights you can self-host - though raw benchmark scores aren't the whole story, as latency, tool-use accuracy, and your task mix will matter as much.
The multimodal angle is the more interesting differentiator. Most of the open-weight coding field is text-only. M3 combines frontier-level coding performance with a 1M-token context window and native multimodal capabilities - image and video understanding - in a single architecture built for long-horizon, complex coding and agentic tasks. If your agent pipeline needs to look at screenshots, diagrams, or video frames and then write code, the field of open-weight options gets thin quickly. M3 is one of the very few places to start.
The risks a launch post won't tell you
There are two things worth naming plainly before you build on M3.
First, sourcing. China's 2017 National Intelligence Law requires MiniMax to support, assist, and cooperate with Chinese government intelligence work - an obligation that applies to every prompt processed through MiniMax's API endpoint, regardless of where the user is located. This is not a reason to rule out M3; it is a reason to route sensitive workloads through the self-hosted weights rather than the API, and to have that conversation with your security team before you ship.
Second, the weights are now public, but speed claims come from MiniMax's own benchmarking, and the reproducible implementation details that would let outside engineers verify the numbers under independent conditions had not been published at launch. Community replication is underway. Until independent results on Terminal-Bench and long-horizon agentic tasks accumulate, treat the MSA throughput figures as directional.
Every model in this class performs significantly better inside a structured agent harness than in raw chat mode - that investment isn't optional for production use. Benchmark scores are useful for filtering; real-task evaluation on your own codebase is required before committing.
An AI teammate like Beagle that lives in Slack or Teams won't be routing code review through a self-hosted 428B model directly - but the teams building the underlying agentic workflows that feed into those loops absolutely need a clear-eyed view of what M3 is and isn't before they wire it in.
The open-weight frontier coding model landscape a year ago was genuinely thin. MiniMax M3 is a real addition to a field that is now legitimately crowded. That is the honest news.