--- title: "MiniMax M3 Packs Frontier Coding, 1M Context and Vision Into One Open Model" date: 2026-10-18 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-minimax-m3 language: en --- # MiniMax M3 Packs Frontier Coding, 1M Context and Vision Into One Open Model > MiniMax M3 combines open weights, a million-token window, and native image-and-video understanding — the first open model the lab says delivers all three at once. Closed frontier models tend to show off three things at once: strong coding, a huge context window, and the ability to look at images and video. Open-weight models usually pick one or two. MiniMax's pitch with **M3** is that it no longer makes you choose. Released on **June 1, 2026**, M3 is the company's latest open-weight foundation model, and MiniMax frames it as the first and only open model to bring frontier-level coding, a **1M-token context window**, and native multimodality into a single architecture. ## The attention trick behind the long context The enabling technology is **MiniMax Sparse Attention (MSA)**, a sparse-attention architecture the team proposed. Standard full attention scores every token pair, which grows quadratically with context length — at a million tokens, that becomes computationally brutal. MSA adds a lightweight pre-filtering stage that selects the relevant blocks of the key-value cache, then runs the main attention only on those blocks. MiniMax claims concrete gains from this. At a one-million-token context, per-token compute drops to about **one-twentieth** of the previous-generation model, with more than **9x faster prefill** and more than **15x faster decoding**. The team also says MSA matches full attention on the vast majority of capabilities across multiple ablations, and that it beats open alternatives like Flash-Sparse-Attention and flash-moba by more than 4x on arithmetic intensity under M3's head configuration. ## Coding and agents are the headline MiniMax centers M3 on software engineering and agentic work. On the company's own benchmark table, the model posts **59.0% on SWE-Bench Pro**, **66.0% on Terminal-Bench 2.1**, **34.8% on SWE-fficiency**, **28.8% on KernelBench Hard**, and **74.2% on MCP Atlas**. The lab argues these reflect real engineering: M3 was trained and evaluated with an interactive user simulator that mimics how developers actually collaborate — clarifying requirements, correcting course, and iterating across a session. To show agentic endurance rather than a curated score, MiniMax says it tasked M3 with independently reproducing an ICLR 2025 outstanding paper's experiments; the model ran for nearly 12 hours, produced 18 code commits, and generated 23 charts without human intervention. That is a single demonstrative run, not a controlled benchmark, but it is a more honest test of long-horizon agency than most leaderboard rows. ## Native multimodality, including computer use M3 is trained with mixed modalities from step zero, so it accepts image and video input and can operate a desktop computer — "computer use" in the sense of driving a GUI to complete tasks. MiniMax contends that interleaved training lets the semantic spaces of different modalities merge more naturally than synthetic bridging does, and that interleaved data scales more easily than purely synthetic data. This is the capability that, combined with the 1M window and coding strength, defines M3's positioning: a single open model that can read your screen, remember a long session, and write the code to act on it. ## Pricing and availability M3 is available through MiniMax Code, a token plan, and the company's API. Secondary model trackers list an API list price of about **US$0.60 per million input tokens** and **US$2.40 per million output tokens**, with a standing discount bringing it to roughly **US$0.30 / US$1.20** — a small fraction of US closed-frontier rates. Those price points come from third-party trackers rather than MiniMax's primary launch post and should be confirmed in the live console. On parameters, third-party spec sheets describe M3 as a large sparse mixture-of-experts with roughly **428 billion total parameters** and about **23 billion active per token**, released under MiniMax's own community license with commercial restrictions rather than a fully permissive Apache/MIT term. Weights were reported on Hugging Face shortly after the API launch. ## Why "open-weight" still has fine print M3 is open-weight, but not as unrestricted as some peers. MiniMax's community license typically gates high-revenue commercial use, so enterprises should read the license before shipping a product on top of the model. The trade-off is real: you get frontier-adjacent capability and self-hosting, but not necessarily the unrestricted freedom of an Apache-2.0 checkpoint. ## How it compares to the China open-wave M3 sits alongside a crowded 2025–2026 field of Chinese open-weight releases — DeepSeek's sparse-attention experiments, Alibaba's multimodal Qwen3.5, Tencent's hybrid Mamba reasoning model, and others. MiniMax's distinct claim is the combination: long context plus coding plus vision in one downloadable model, rather than excellence in any single axis. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, M3's importance is that it treats a million-token window and native vision as baseline table stakes for an open model, which raises the floor for what self-hosted agents can do without calling a closed API. ## What readers can do now - If your agent needs long memory plus screen understanding, test M3's 1M-token window and computer-use on a real multi-step task rather than a leaderboard snippet. - Before self-hosting, read MiniMax's community license — the commercial-use gate differs from Apache-2.0 models and affects what you can ship. - Compare M3's API cost against a closed frontier model on the same coding workload; the per-token gap is the real story for budget-conscious teams. ## Honest limitations Primary facts — the June 1, 2026 release, the MSA architecture, the 1M-token window, native image/video input with computer use, and the SWE-Bench Pro / Terminal-Bench / MCP Atlas figures — come from MiniMax's official launch post. Those benchmark numbers are vendor-reported on the company's own table, not independently audited here. Parameter counts (~428B total / ~23B active) and API pricing (US$0.60 / US$2.40 list, discounted to ~US$0.30 / US$1.20) are drawn from third-party model trackers and were not verified against MiniMax's primary materials. We did not run M3 ourselves, so real-world latency, multimodal quality, and cost should be measured on your own hardware and workloads. License terms and high-revenue commercial restrictions may change and must be confirmed directly with MiniMax. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-minimax-m3 Free to quote with attribution and a link to the original.