Industry Signals from Two Open-Weight Models

entrup · x · 2026-07-18

This post interprets two recent "open model" announcements: First, Mira Murati's Thinking Machines Lab reportedly released Inkling, a 975B-parameter multimodal open-weight model supporting text/image/audio inputs with up to a 1M token context, tailored for enterprise on-premises deployment and fine-tuning. Second, Moonshot AI released Kimi K3, described as a 2.8T-parameter sparse MoE model, also supporting a 1M context and native vision, excelling in coding, long-range reasoning, and agent workflows, with plans to open-source the weights before July 27.

The author's core thesis is that frontier models are shifting from closed APIs toward large-scale open weights, and enterprise demand for data/IP control will drive "sovereign AI" solutions. Meanwhile, US-China lab competition is rapidly narrowing the gap between closed and open source, pushing future models toward multimodal, long-context, and agentic practical scenarios.

Original post →

More from Companies & People

Companies & People channel →