Trinity Large Preview

Trinity Mini is a 26B-parameter (3B active) Trinity MoE model with 128K context, tuned for strong reasoning, function calling and multi-step agents while remaining efficient for enterprise backends.

Trinity Nano Base is the base 6B (1B active) Trinity Nano checkpoint, pre–fine-tuning, meant to be domain-tuned rather than used directly for chatting, trained on about 10T tokens under Apache-2.0.

Brumby-14B-Base is Manifest AI’s 14B attention-free language model, a retrained Qwen3-14B variant that replaces attention with Power Retention layers for hardware-efficient, long-context reasoning under Apache 2.0.

Qwen 2.5-VL-72B is Alibaba’s flagship open-weight vision-language model. It takes images (docs, charts, screenshots, photos) plus text and answers in text, with strong OCR, layout understanding, and multi-image reasoning. It supports long context, function/tool calling, and reliable JSON outputs—ideal for multimodal RAG, agents, and enterprise workflows.

The MiniMax M2.5 announcement page is dated “2026.2.12” and opens with “Today we’re introducing our latest model, MiniMax-M2.5,” indicating a release on 12 February 2026.