CONFIRMEDDate confidence: CONFIRMEDStated as firmHow firmly the source words the date: confirmed, scheduled, estimated, delayed or unverified.Show all CONFIRMED pins86% CONFIDENCEOverall confidence: 86%How well the pin's source and references back up its dates.Weighted average of how firmly 6 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better
MiniMax's[1][4] post is dated 2026-06-01 and begins "MiniMax M3 is officially released today"; its API platform release notes list the model on Jun. 1, 2026.- Start
- Jun 1, 202690% CONFIDENCEfrom the source
MiniMax-M3 Released
MiniMax releases MiniMax-M3, a natively multimodal model with a 1M-token context built on its new sparse attention, scoring 59.0% on SWE-Bench Pro and priced at $0.30 per 1 million input tokens after a launch discount.
- MiniMax released MiniMax-M3 on 1 June 2026, calling it "currently the first and only open-weight model" to combine frontier coding and agent performance, a 1M-token context and native multimodality with desktop computer use[1]
- The enabler is MiniMax Sparse Attention (MSA), a new attention design that MiniMax says cuts per-token compute at 1M context to 1/20 of the previous generation, with more than 9x faster prefill and 15x faster decoding[1][5]
- MiniMax's coding scores are 59.0% on SWE-Bench Pro (above GPT 5.5's 58.6 and Gemini 3.1 Pro's 54.2, below Claude Opus 4.7's 64.3), 66.0% on Terminal Bench 2.1, 28.8% on KernelBench Hard and 74.2% on MCP Atlas[1]
- It launched first on MiniMax's API at $0.60 per 1 million input tokens and $2.40 for output with 50% off for seven days, which is the $0.30 and $1.20 that MiniMax now lists as a permanent discount and that Artificial Analysis measures[1][3][4]
- Weights followed within days on Hugging Face under the MiniMax community licence at about 428B total and 23B active parameters[2], against 229.9B and 9.8B active and a 192K context for the M2 series[6]; MiniMax had teased M3 on 27 May[6]
- Artificial Analysis scores it 29 on its Intelligence Index, 18th of 117 models, at about 107 output tokens per second[3]
Notable features
- Multimodal from step zero: trained on mixed text, image and video data, it takes image and video input and can operate a desktop; thinking can be switched on or off at the same price[1][2]
- Long-horizon demos: it reproduced an ICLR 2025 outstanding paper on LLM fine-tuning dynamics in about 12 hours (18 commits, 23 figures), and lifted an FP8 matrix-multiplication kernel on Hopper GPUs from 7.6% to 71.3% of peak over 24 hours and 147 submissions (9.4x)[1]
- PostTrainBench: it scored 0.37 at autonomously post-training four base models in 12 hours, against 0.42 for Claude Opus 4.7 and 0.39 for GPT 5.5[1]
- MiniMax Code, an agent trained alongside M3 with Agent Team workflows and computer use, built on the OpenCode and Pi harnesses; Token Plans of $20 (about 1.7B tokens a month), $50 (5.1B) and $120 (9.8B)[1]
- API price steps up above 512k input tokens to $1.20 per 1 million input and $4.80 output at list, and a priority service tier is available through sales support[1][4]
Benchmarks[1]
| Benchmark | MiniMax M3 | Claude Opus 4.7 | GPT 5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-Bench Pro | 59.0 | 64.3 | 58.6 | 54.2 |
| Terminal Bench 2.1 | 66.0 | 66.1 | 78.2 | 70.0 |
| VIBE V2 | 50.1 | 55.8 | 50.5 | 28 |
| SVG-Bench | 63.7 | 62.3 | 58.2 | 59.2 |
| KernelBench Hard | 28.8 | 30.7 | 20.9 | 18.6 |
| BrowseComp | 83.5 | 79.3 | 84.4 | 85.9 |
| GDPval rubrics | 74.7 | 79.8 | 80.6 | 57.8 |
| BankerToolBench | 76.1 | 81.3 | 70.0 | 67.0 |
| MCP Atlas | 74.2 | 77 | 75.3 | 69.2 |
| OSWorld-verified | 75.2 | 82.8 | 78.7 | 76.2 |
- Read from MiniMax's benchmark chart on the launch post; the post also lists SWE-fficiency 34.8[1]