MiniMax-M2.7

MiniMax announced M2.7 as the first model in its line that deeply participated in its own evolution, using agent teams, complex skills, and dynamic tool search to update its own memory and build dozens of harness skills for reinforcement learning experiments. The release emphasizes self-evolution, where M2.7 improves its learning process and harness based on experiment results. This positions the model for highly elaborate productivity tasks beyond earlier M2-series iterations. Benchmark results reported by MiniMax include 56.22% on SWE-Pro, 55.6% on VIBE-Pro, and 57.0% on Terminal Bench 2, alongside a GDPval-AA ELO of 1495. The model shows improved complex editing in Office suite applications including Excel, PowerPoint, and Word, with better handling of multi-turn modifications and high-fidelity edits. M2.7 also demonstrates stronger identity preservation and emotional intelligence, expanding its use into broader interactive entertainment scenarios.

OpenRouter's listing explicitly names MiniMax-M2.7 with a release date of Mar 18, 2026, a 205K-token context window, and headline benchmarks: 56.2% on SWE-Pro, 57.0% on Terminal Bench 2, and a 1495 ELO on GDPval-AA, framed as a next-generation LLM designed for autonomous, real-world productivity and continuous improvem The same page positions M2.7 for production-grade workflows including live debugging, root cause analysis, financial modeling, and full document generation across Word, Excel, and PowerPoint. Pricing metadata is shown ($0.24/$0.96 per 1M tokens at a 60% promotional rate, plus a weighted-average effective input of $0.13

Opper AI's release tracker explicitly names MiniMax M2.7 as a dated entry in the MiniMax M-series release history, placing its launch on 18 March 2026 with a 197K context window, $0.25/$1.00 per million input/output token pricing, and an Intelligence score of 39 (per Artificial Analysis, cross-checked against vendor an The M2.7 entry on Opper's tracker is a single line among seven MiniMax releases, so it offers thin technical substance: no architecture details, no benchmark breakdown, and no developer-facing guidance. Pricing and intelligence figures are sourced from Artificial Analysis rather than MiniMax's own documentation, so the

MiniMax M2.7 is a sparse mixture-of-experts LLM featuring 230B total parameters with 10B active per token at a 4.3% activation rate, supported by a 200K input context window. The architecture uses 256 local experts with 8 activated per token across 62 layers, paired with multi-head causal self-attention using Rotary Position Embeddings and QK RMS Normalization during training. The release adds enhancements to the earlier MiniMax M2.5 model for agentic harnesses and complex workloads including reasoning and engineering tasks. NVIDIA and open-source contributors integrated QK RMS Norm and FP8 MoE optimizations into vLLM and SGLang, delivering up to 2.5x throughput with vLLM and 2.7x with SGLang on Blackwell Ultra GPUs. Developers can deploy M2.7 via vLLM with tensor and expert parallelism, or via SGLang with FP8 quantization and FlashInfer kernels. Post-training is available through NVIDIA NeMo AutoModel and NeMo RL with GRPO configurations for fine-tuning and reinforcement learning workflows.

MiniMax introduced M2.7 as an LLM focused on building complex agent harnesses for highly elaborate productivity tasks. The company highlights strong real-world software engineering performance, including end-to-end project delivery, log analysis for bug hunting, code security, and machine learning tasks. On GDPval-AA, M2.7 achieves an ELO score of 1495, which MiniMax calls the highest among open-source models at release. On benchmarks, M2.7 scores 56.22% on SWE-Pro, approaching the best Opus level, with 55.6% on VIBE-Pro for end-to-end project delivery and 57.0% on Terminal Bench 2 for complex engineering systems. The model maintains a 97% skill adherence rate across 40 complex skills of about 2000 tokens each, and shows significant gains over M2.5 on MMClaw evaluation, approaching the level of the latest Sonnet 4.6.

InferenceX publishes a third-party technical profile of the MiniMax-M2 series that explicitly names MiniMax-M2.7 alongside M2.5. The page documents M2.7's 2026-03-18 release, the dual serving variants "MiniMax-M2.7" and "M2.7-highspeed," and the underlying MoE architecture (230B total parameters, ~10B active, 62 layers The same profile reports an internal M2.7 version that autonomously optimized a programming scenario and surfaces positioning focused on agentic tool use, coding, and productivity workflows. The page is not itself a MiniMax announcement but cites MiniMax's M2.7 announcement page and API release notes for the 2026-03-18

The Together AI model listing for MiniMax M2.7 restates concrete benchmark numbers and capability claims attributed to the model itself, independent of any host-specific SLA or availability marketing. It reports SWE-Pro at 56.22% (matching GPT-5.3-Codex), 55.6% on VIBE-Pro (near Opus 4.6) for end-to-end Web/Android/iOS The same page describes M2.7's self-evolution workflow: an internal version ran 100+ autonomous optimization rounds — analyzing failure trajectories, modifying code, evaluating results, and deciding to keep or revert — to achieve a 30% improvement on internal programming benchmarks, while during training the model upda

A third-party deep-dive on MiniMax M2.7, updated March 2026, consolidates figures from the official March 18, 2026 technical report and frames M2.7 as a self-improving model for complex agent harnesses, tool-chain coordination, and multi-stage productivity tasks. The page reports concrete engineering-benchmark scores: The same deep-dive emphasizes that M2.7's core value is delivery reliability under messy production constraints, not just stronger coding output, making it useful as a continuous optimization component in model and harness development operations. For engineering managers evaluating the model, the aggregated benchmark s