Inkling is a general-purpose multimodal model released by Thinking Machines Lab with open weights under the Apache 2.0 license. It is designed to handle text, image, and audio inputs and generate text outputs, making it suitable for conversational AI, instruction-following, coding assistants, agentic systems, retrieval-augmented generation, and other multimodal applications. The model supports English and has broader multilingual capabilities, as well as use across multiple coding languages.
Technically, Inkling is a 66-layer decoder-only autoregressive transformer with a sparse Mixture-of-Experts architecture. Each token is routed to 6 of 256 experts, with 2 shared experts always active. It uses hybrid local and global attention, and processes images and video through a hierarchical patch encoder while audio is encoded as discrete tokens in a shared hidden space. The model has 975 billion total parameters, with 41 billion active at inference time, and supports BF16 and NVFP4 numerics.
Inkling accepts UTF-8 text, pixel-based images, and 16 kHz WAV audio, with optimal image sizes between 40 and 4096 pixels per dimension and audio under 20 minutes. Training data came from public internet sources, public repositories, third-party data, and synthetic or augmented data, with cleaning, deduplication, and filtering applied for quality and safety.
Evaluation results show strong performance on reasoning, math, knowledge, and coding benchmarks, including HLE, AIME 2026, GPQA Diamond, and SWE-bench Verified. The model is available through Tinker, local deployment frameworks such as SGLang and vLLM, and third-party inference providers.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Comments
Login to view and post comments.