MiniMax M2.5/M2.7 ships
MiniMax released M2.5 and M2.7 models that scored top marks on SWE‑bench (80.2% verified) and claim low input costs (~$0.19 per million tokens), with TEE privacy options on platforms like Chutes. The models are being positioned as cheaper, high‑quality options for coding and agentic workflows. (x.com)
MiniMax published the M2.5 release on February 12, 2026 and uploaded model artifacts and a model card to Hugging Face while maintaining a companion repository on GitHub. ( ) The company documents an M2.5-Lightning variant engineered for throughput around 100 tokens per second and says that variant completes complex agentic tasks markedly faster than earlier M2-series builds. (minimax.io) Pricing trackers and API directories report provider-dependent input pricing for M2.5 that varies by speed tier and reseller, with some listings showing per‑million-token input prices in the mid‑teens to low‑tenths of a dollar and higher output-token rates. ( ) MiniMax announced M2.7 on March 18, 2026 as a “self‑evolving” generation that the company says ran 100+ autonomous optimization loops and can automate roughly 30–50% of certain reinforcement‑learning research workflows in their internal pipeline. ( ) Early analyses and vendor write‑ups credit M2.7 with lower hallucination rates and measurable gains on multi‑tool, agentic benchmarks relative to M2.5, and both models are listed with context windows of roughly 196K–205K tokens in public model cards and third‑party summaries. ( ) Third‑party hosting and tooling integrations are already visible: MiniMax images and endpoints appear on Ollama and OpenRouter, and Chutes has published a dedicated MiniMax TEE chute indicating a confidential‑compute deployment option. ( ) Chutes’ documentation and blog posts describe hardware-backed Trusted Execution Environments, end‑to‑end encryption, and remote‑attestation endpoints for verifying TEE integrity, and Chutes’ privacy page states minimal metadata retention for TEE inferences. ( )
