MiniMax-M3

A release tracker confirms MiniMax M3 was released on June 1, 2026, as the newest entry in the MiniMax model lineup. The tracker lists M3 with a 512K context window, input pricing of $0.30 per million tokens, and output pricing of $1.20 per million tokens, with an intelligence score of 45. This date corroborates the of The tracker lists six prior MiniMax releases dating back to October 2025, showing the progression of the M-series from M1 through M2 and its variants to M3. The 512K context figure shown here differs from the 1M context reported by the vendor and other third-party analyses, likely reflecting a served variant or aggrega

An original third-party experiment sent 300 paid requests to the public MiniMax OpenAI-compatible Chat Completions endpoint, scoring each returned function name and arguments against deterministic ground truth across 100 English prompts and 10 JSON Schema families. MiniMax M3 produced a strict pass in 298 of 300 trials The benchmark makes a useful distinction between API-level success (100%) and application-level correctness, with one response truncated at the output limit producing incomplete JSON arguments and one returning ordinary assistant text without a tool_calls entry. The complete dataset, request fixtures, raw JSONL and CSV

An independent hands-on coding-agent benchmark evaluated MiniMax-M3 on ten planted-bug JavaScript and Python repositories using a tool harness with list_files, read_file, write_file, and related operations, exposing only public tests and grading against hidden assertions afterward. On the primary result, M3 resolved 7 A noteworthy finding for developers is that the three failures were not capability-driven: the model consumed the entire 4,096-token completion budget during reasoning and never invoked the write tool, making completion-budget and harness design material factors in real coding performance. Complete task definitions, sa

A third-party technical breakdown confirms MiniMax M3 as a 428-billion-parameter mixture-of-experts model with approximately 23 billion active parameters per token, released June 1, 2026. The analysis details that the attention stack combines grouped-query attention with MiniMax Sparse Attention (MSA), which reportedly The breakdown reports that model weights went live on Hugging Face by June 7, 2026, and the technical report was published on arXiv on June 11, 2026, under a custom MiniMax-community license. It also notes an interleaved-thinking tool-calling behavior as a practical integration consideration. Official pricing is listed

An independent technical deep-dive (July 5, 2026) details the architecture behind MiniMax-M3, describing it as a Mixture-of-Experts model with roughly 428B total parameters and about 23B active per token, paired with MiniMax Sparse Attention (MSA). MSA is explained as a two-branch design: a lightweight Index Branch tha Beyond architecture, the piece contrasts benchmark reality against the marketing claims and walks through production integration via an OpenAI-compatible unified gateway. For developers evaluating M3, the core takeaway is that long-context efficiency is real and hardware-dependent (best on H800-class GPUs), while the g

MiniMax M3 is an open-weight, multimodal Mixture-of-Experts model released on June 1, 2026, built for coding, tool use, long-running agents, and large-context workloads. According to the model profile, it has approximately 428 billion total parameters with about 23 billion active per token, accepts text, image, and vid Standard-tier API pricing is $0.30 per million input tokens and $1.20 per million output tokens up to 512K context, doubling to $0.60/$2.40 per million for the 512K–1M band, and the profile notes full-scale self-hosting remains a data-center-class undertaking while headline benchmark scores still come largely from vend

A Verdent developer guide dated June 15, 2026, cross-referenced against MiniMax's official model card and the arXiv technical report (arXiv:2606.13392), provides builder-oriented architectural detail on MiniMax M3. The page describes M3 as a 428B-parameter Mixture-of-Experts model with approximately 22B parameters acti The guide flags a non-standard license as a caveat often skipped in other coverage, and emphasizes that M3's headline metrics — 1M context, multimodality, and open weights — need to be translated into engineering decisions around long-repo reading, tool-calling loops, and multi-step agentic tasks. It positions MiniMax'

An independent benchmark comparison (ofox.ai, June 15, 2026) reports MiniMax-M3 scoring 59.0% on SWE-Bench Pro versus GPT-5.5's 58.6%, with M3 priced at $0.60 input / $2.40 output per million tokens compared to GPT-5.5's $5 / $30 — roughly 8–12× cheaper per SWE-Bench Pro point. The post frames M3 as the first open-weig The comparison notes GPT-5.5 retains advantages on Terminal-Bench (82.7% vs M3's 66.0%) and Codex CLI integration, while M3 is favored for long-context refactors (1M context with MSA, ~15× faster decode than M2.5), native multimodal code review, and air-gapped on-prem deployment via open weights on Hugging Face.

Fireworks AI published a launch post (June 12, 2026) announcing Day-0 support for MiniMax-M3, supporting up to 500K tokens of context at launch and partnering with MiniMax to bring the full 1M-token window shortly after. The post frames M3 as MiniMax's flagship frontier model — a 500K-token context open-weight release The post cites Artificial Analysis' third-party intelligence index, stating M3 surpasses all other open-source models in overall intelligence and exceeds several closed-source models including Opus 4.6. (Note: this candidate is largely a serving-provider announcement, so it is down-weighted but still useful to corrobor

Artificial Analysis published an independent third-party evaluation on June 8, 2026, scoring MiniMax-M3 at 55 on the Artificial Analysis Intelligence Index, placing it ahead of open-weights peers Kimi K2.6 (54) and MiMo-V2.5-Pro (54). M3 is MiniMax's first multimodal M-series model, adding image/video input and a 1M-to Specific benchmark improvements over M2.7 include HLE +9 (28%→37%), GPQA Diamond +6 (87%→93%), AA-LCR +5 (69%→74%), IFBench +7 (76%→83%), and CritPt +3 (1%→4%), with a small SciCode regression (47%→45%). M3 scores 1670 on GDPval-AA (behind Claude Opus 4.8 max at 1890 and GPT-5.5 xhigh at 1769, level with Claude Sonnet

Navneet Guglani's Medium analysis provides third-party technical context for MiniMax-M3, confirming its June 1, 2026 launch and its status as the first open-weight model to combine frontier-level coding, a 1-million-token context window, and native multimodal input in one architecture. The piece traces MiniMax's rapid Beyond launch details, the article notes that the M3 API went live on the same day as the release, with open weights and a technical report scheduled within 10 days. It positions M3 against closed-source frontier models such as Claude Opus, GPT-5.5, and Gemini 3.1 Pro, arguing that while open models have previously exc

An independent Stackademic hands-on test (June 3, 2026) puts MiniMax M3 through a full day of real Python and agentic workloads against Claude Opus 4.8 and GPT-5.5, using identical prompts, codebases, and workflows. The reviewer reports M3 scoring roughly 59% on SWE-Bench Pro (close to Claude Opus 4.8), strong performa The write-up describes M3 as an open-weights model built specifically for coding, long-context work, and AI agents, with its sparse attention mechanism enabling the 1M-token context window by focusing compute on the most relevant prompt regions rather than every token. While useful as a third-party developer signal on