MiniMax M3 announcement — first open-weights model combining coding, agentic, and natively-multimodal frontier capabilities

TL;DR

MiniMax announced M3 on Jun 1 2026, positioning it as “the first open-weights model to combine three frontier capabilities”: coding/agentic, 1M-context efficiency via MiniMax Sparse Attention (MSA), and native multimodality from step zero. Headline benchmarks: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas. The API and MiniMax Code product are live now; open weights and the tech report are promised in ~10 days. No architecture details, parameter count, or active-parameter ratio are disclosed in the announcement.

Key claims

Combines coding/agentic, long-context, and multimodal frontiers in a single open-weights model — framed as a first [tweet body].
Coding/agentic numbers: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas [tweet body].
“MiniMax Sparse Attention scales context to 1M” — branded variant of sparse attention, presumably the M-series successor to M1’s lightning-attention hybrid and M2’s dense softmax [tweet body].
“Natively Multimodal from Step Zero” — distinguishing it from M1/M2/M2.5/M2.7 which were text-only [tweet body].
Weights and technical report ETA ~10 days from announcement [tweet body].

Method

Not disclosed in the announcement tweet. Inferences worth flagging when the tech report drops: (a) whether MSA is a new architecture or a rebranding of the lightning/softmax hybrid used in MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention or something closer to DeepSeek Sparse Attention; (b) the active/total parameter split (M2/M2.5/M2.7 ran 230B-total / 10B-active); (c) what “natively multimodal from step zero” means architecturally — joint pretraining vs. late-fusion adapters.

Results

Benchmark numbers as stated in the tweet, no comparison context provided:

Benchmark	M3
SWE-Bench Pro	59.0%
Terminal Bench 2.1	66.0%
SWE-fficiency	34.8%
KernelBench Hard	28.8%
MCP Atlas	74.2%

For SWE-Bench Verified context, the M2 page recorded 69.4 and the cluster on Agentic Software Engineering tracks competing open models. M3’s SWE-Bench Pro score is on the harder Pro split, so it isn’t directly comparable to M2’s Verified number without re-running both.

Why it’s interesting

Continues MiniMax’s monthly-cadence open-weights drops (MiniMax-M2: a Mini model built for Max coding & agentic workflows → MiniMax-M2.5 → MiniMax M2.7: Early Echoes of Self-Evolution → M3), now extending the line into multimodality for the first time.
MSA branding is the headline architectural claim. If MSA is genuinely a new sparse-attention design rather than M1’s lightning-attention hybrid, M3 joins the cluster on Hybrid Linear Attention alongside DeepSeek-V3.2’s DSA (DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention), Kimi Linear (Kimi Linear: An Expressive, Efficient Attention Architecture), and Qwen3-Next as the latest answers to scaling context past 128K.
“Natively multimodal” framing echoes the line from Future of multimodal lies in native pixel-space unified architectures (Yuwei Niu quote-RT of NEO-Unify) and Emu3.5: Native Multimodal Models are World Learners that pixel-/token-unified-from-pretraining is the next architectural watershed.
Pre-release announcement (weights in ~10 days) — this page should be revisited and the architecture section filled in once the tech report lands.