975B total / 41B active (MoE)
1M tokens
Apache 2.0
Official open-weight flagship release from Thinking Machines Lab positioned as a broad multimodal base model for customization, fine-tuning, and agentic product workflows rather than a narrowly benchmark-optimized frontier model. The official announcement and model card describe Inkling as a 66-layer decoder-only multimodal autoregressive transformer with a sparse Mixture-of-Experts feed-forward backbone, routing each token to 6 of 256 experts plus 2 shared experts, with hybrid local/global attention and native joint processing of text, image, audio, and video representations in a shared hidden space. Thinking Machines says the model has 975B total parameters with 41B active parameters, supports up to 1M tokens of context, and was pretrained from scratch on 45T tokens spanning text, images, audio, and video. The release announcement also introduces Inkling-Small as a lighter-weight 12B-active preview trained with a similar recipe, although this entry covers the publicly verifiable main Inkling release. Public materials emphasize controllable thinking effort, strong agentic coding and tool-use behavior across varied harnesses, broad multimodal capability, and compatibility with Tinker for fine-tuning and self-customization workflows, including a demo where Inkling fine-tunes and reloads its own weights. The public benchmark tables position Inkling as a balanced generalist open-weight model with competitive results across reasoning, coding, agentic, factuality, chat, vision, audio, and safety evaluations, including HLE, AIME 2026, GPQA Diamond, SWE-Bench Verified, SWE-Bench Pro Public, Terminal Bench 2.1, MCP Atlas, BrowseComp, Global-MMLU-Lite, MMMU Pro, CharXiv RQ, Audio MC, and VoiceBench. The model card states that the model supports BF16, MXFP8, and NVFP4 numerics, while the public Hugging Face release exposes BF16 and NVFP4 checkpoint formats for local deployment through SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face tooling.