Thinking Machines Lab Ships Inkling: 975B-A41B Multimodal, Apache 2.0 — the Leading American Open-Weights Model

Latent Space (AINews) / Thinking Machines Lab·high signal

TML's first full LLM is open-weights: 975B total / 41B active MoE, multimodal (text/image/audio in, text out), trained on 45 trillion tokens, with a 1M-token context in the open weights (256K via API/Tinker). A smaller Inkling-Small (276B-A12B) ships alongside. It ranks #41 on the Intelligence Index — top U.S. open-weights model but behind GLM-5.2 and Kimi K2.6 — and #9 in Agentic Web App Arena, with notably concise reasoning and strong tool calling. Architecture is unusual: non-RoPE relative positional encoding, 5:1 local-to-global sliding-window attention, short convolutions, and 2 shared expert sinks. Available on Hugging Face, Databricks, Baseten, Modal, and vLLM.