Quiet Wave of New Models
A flurry of new model releases surfaced this week—edge‑friendly Kimi K2.5, reasoning‑focused MiroThinker 72B, iPhone‑capable Qwen 3.5 and Google’s Nano‑Banana 2 (notable for sub‑second 4K image synthesis). Google is also doubling down on Gemini as a platform play, signaling startups will face both more model choice and tighter platform competition. (labla.org; fastcompany.com)
Moonshot's Kimi K2.5 was pretrained on roughly 15 trillion mixed visual and text tokens and is published as an open-source native multimodal, agentic model on Hugging Face and GitHub. (huggingface.co) MiroMind's MiroThinker appears as a 72‑billion‑parameter research agent on Hugging Face and is explicitly built for interactive scaling to support deeper agent–environment loops. (huggingface.co) The team's arXiv paper documents the "interactive scaling" training regimen designed to improve iterative tool use and stepwise verification in complex reasoning tasks. (arxiv.org) Alibaba's Qwen 3.5 release includes an open‑weight flagship (Qwen3.5‑397B) alongside smaller, deployable variants intended for on‑device use. (qwen.ai) Community demos and vendor write‑ups show 4‑bit quantized Qwen 3.5 variants running on iPhone hardware (examples cited on iPhone 15/17 families) via MLX/optimized runtimes at 0.8–9B sizes. (techeblog.com) Google's Nano Banana 2 is the Gemini 3.1 "Flash Image" configuration optimized for sub‑second, studio‑quality 4K image synthesis and improved subject consistency. (blog.google) Google has pushed Gemini as a startup platform through the "Gemini Kit" and related programs, and a public cohort selection that included 53 startups from nearly 1,000 applicants. (aragonresearch.com) Gemini's technical push (integration across Search, Android, Workspace and Vertex AI) combined with product rollouts like Nano Banana 2 (Feb 26, 2026) tightens the commercial calculus for early‑stage companies weighing whether to build on Google’s stack or on alternative open models. (blog.google)
