Inkling, an open-weights model
On 15 July 2026 Thinking Machines Lab released Inkling, a Mixture-of-Experts model of 975 billion parameters (41 billion active) with its full weights on Hugging Face. It reads text, images and audio and writes text, takes a context of up to 1 million tokens, and was pretrained on 45 trillion tokens. The lab says it is the first model of its family and not the strongest model available, closed or open, but a base to customise.
Why it matters
A lab known for a fine-tuning service releases a large base model under a permissive licence; by its own account the aim is customisation, not leading rankings. It puts the open weights of a model of this size from a new laboratory on the record.
What the post says. A Mixture-of-Experts transformer: 975B parameters in total, 41B active; a context of up to 1M tokens; pretraining on 45 trillion tokens of text, images, audio and video; text, image and audio input, text output; controllable thinking effort. 'The first in a family of models of different sizes'. Available for fine-tuning on the Tinker platform the same day (at a 50% limited-time discount) and through APIs on Together AI, Fireworks, Modal, Databricks and Baseten; supported in SGLang, vLLM, llama.cpp and transformers. Full weights are on Hugging Face, as the original checkpoint and as an NVFP4 checkpoint for NVIDIA Blackwell. Weights and licence. The Hugging Face API record for the repository thinkingmachines/Inkling (read on 10 October 2026): not gated, licence apache-2.0, 108 safetensors files, created on 14 July 2026 at 13:23 UTC. The record gives the day of the post, 15 July; the model card was not read. The lab's figures. SWE-bench Verified 77.6% (with a bash-only harness), AIME 2026 97.1%, GPQA Diamond 87.2%, HLE text-only 29.7%. The post notes that the HLE curve in its efficiency chart reflects an earlier checkpoint that scores slightly below the final release. The lab's table sets beside them Claude Fable 5 (SWE-bench Verified 95.0%) and GPT 5.6 Sol (82.2%); the post admits that some Terminal Bench solutions were found contaminated by web search and scored 0. Safety was evaluated with internal tests and external testers. Inkling-Small. In the same post the lab showed a preview of Inkling-Small (276B parameters, 12B active) and promised weights after testing. Full weights were released with a post of 30 July (the Hugging Face repository was created on 27 July at 22:33 UTC); there is no record of its own for it here. What the record does not claim. All figures are the lab's; no independent evaluations were read. The post says 'first in a family', not 'the first open-weights model in the world'.