MiniMax: MiniMax M3

MiniMax-M3 is the Shanghai-based lab's flagship model, officially released on June 1, 2026. It's positioned as the first open-weight model to combine three frontier capabilities in one architecture: strong coding/agentic performance, a 1-million-token context window, and native multimodality (text, image, and video understanding).

Highlights:

• Native Multimodality: M3 undergoes mixed-modality training from the very first step, enabling deeper semantic fusion across text, image, and video.
• Context Scaling via Sparse Attention: M3 introduces MiniMax Sparse Attention (MSA) to improve long context efficiency. M3 delivers 9× prefill and 15× decode speedups compared to M2 at 1M context, reducing per-token compute to 1/20.
• Coding & Cowork Capability: M3 achieves frontier-level performance across long-horizon agentic benchmarks, excelling in both coding and cowork.

#Architecture

M3 is built on MiniMax Sparse Attention (MSA), a custom sparse-attention design that replaces full attention with KV-block selection. This is what makes the huge context window practical rather than just a spec on paper — it cuts per-token compute at long context to roughly 1/20th the cost of MiniMax's previous generation, with much faster prefill and decode while largely preserving quality. The model was trained natively as a multimodal system on interleaved text/image/video data, and further tuned for multi-turn, production-style collaboration via an interactive user-simulator framework — oriented toward sustained, multi-step agent work rather than one-shot answers.

Key specs:

• Context window: up to 1,048,576 tokens (1M), with a guaranteed minimum of 512K tokens — about 5x its predecessor, MiniMax M2.7
• Max output: 131,072 tokens
• Precision: FP8
• Modalities: text, image, and video input; text output
• Pricing: roughly $0.30 per million input tokens and $1.20 per million output tokens

#Benchmark Performance

MiniMax reports M3 performing strongly on several coding and agentic benchmarks:

• SWE-Bench Pro: 59.0% — MiniMax's own reporting places this above GPT-5.5 and Gemini 3.1 Pro, though notably behind Claude Opus 4.8's 69.2% on the same benchmark
• Terminal-Bench 2.1: 66.0%
• MCP Atlas: 74.2%
• BrowseComp: 83.5, which MiniMax reports as ahead of Claude Opus 4.7's 79.3
• SWE-fficiency: 34.8%; KernelBench Hard: 28.8%
• On the Artificial Analysis Intelligence Index (a composite reasoning/knowledge/math/coding benchmark), it scores 44 — well above the median of 25 for similarly sized open-weight models

One illustrative capability MiniMax has highlighted: in a roughly 24-hour CUDA kernel optimization task, the model made 147 benchmark submissions using nearly 2,000 tool calls, continuing to improve even after long performance plateaus — offered as evidence of sustained autonomous problem-solving over very long horizons.

#Caveats Worth Knowing

Independent analysts have flagged some reasons for caution around these numbers:

• Several benchmark results were run on MiniMax's own infrastructure, sometimes using agent scaffolding borrowed from other ecosystems (like Claude Code, Mini-SWE-Agent, or Terminus) — which doesn't invalidate the numbers, but means real-world performance inside messy, unfamiliar codebases could differ
• At the time of release, M3 hadn't yet appeared on neutral, third-party boards like DeepSWE, which track long-horizon software engineering tasks independently
• When compared against Anthropic's Opus 4.8 (released around the same time), M3's efficiency-focused architecture showed a real gap on pure code-modification tasks

#Availability and Ecosystem Position

M3 is open-weight and available via API, monthly token plans, and MiniMax's own coding-agent product, with weights and a technical report also distributed through Hugging Face. Within the crowded Chinese open-weight ecosystem, it competes directly with DeepSeek V4 and Alibaba's Qwen3.7-Max — all three targeting similar agentic coding use cases with open licensing, but MiniMax differentiates itself through the specific combination of frontier coding, 1M-token context, and native multimodality in a single model.

#MiniMax Sparse Attention (MSA)

M3 is powered by MiniMax Sparse Attention (MSA), a high-performance sparse attention operator designed for million-token contexts. Compared with GQA, MSA dramatically reduces the attention compute and memory footprint while preserving model quality.

📄 Read the technical report: arXiv:2606.13392 · Hugging Face Papers

#How to Use

• MiniMax Agent
• MiniMax API

M3 supports three reasoning modes through the thinking parameter:

• enabled — Reasoning is always enabled.
• adaptive — M3 automatically determines when additional reasoning is beneficial.
• disabled — Reasoning is disabled to minimize latency and maximize throughput.

#Inference Parameters

We recommend the following parameters for best performance: temperature=1.0, top_p=0.95.

#Contact Us

Contact us at [email protected].