Inkling

What it is

Inkling is the first model released by Thinking Machines Lab. It was published on 15 July 2026 with full open weights on Hugging Face ^[1]. It is a mixture-of-experts (MoE) Transformer, the architecture family in which each layer contains many separate expert subnetworks and only a few of them are used for any given token, so that a very large model can be run at a fraction of the compute.

Architecture

Inkling has 975 billion parameters in total and 41 billion active parameters per token. Each MoE layer holds 256 routed experts plus 2 shared experts, and 6 routed experts are active for each token ^[1]. The context window goes up to 1 million tokens ^[1]. The model reasons natively over text, images and audio, and takes audio and vision input without separate encoders ^[1]. According to Wikipedia, the architecture drew on DeepSeek-V3 ^[2].

Training

Inkling was pretrained on 45 trillion tokens of text, images, audio and video, using NVIDIA GB300 NVL72 systems. Training included large-scale asynchronous reinforcement learning over more than 30 million rollouts ^[1]. For post-training synthetic data it drew on Moonshot AI's Kimi K2.5 ^[2].

Reported results

Thinking Machines reported scores including 87.2% on GPQA Diamond, 97.1% on AIME 2026, 77.6% on SWE-Bench Verified, 46.0% on HLE with tools and 73.5% on MMMU Pro ^[1]. The reported evaluations were run at effort 0.99 and temperature 1.0 ^[1]. The authors described the model plainly, saying it "is not the strongest overall model available today, open or closed" ^[1].

Safety work

Thinking Machines said it trained Inkling to an internal specification of safe model behavior across all modalities ^[1], and that it commissioned external safety testers to verify the results ^[1]. Pre-release testing, including red-teaming by four outside organisations, found no material added risk beyond existing open-weight models ^[3].

Licence and availability

Inkling was released under the Apache licence ^[2]. The full weights on Hugging Face include the original checkpoint and an NVFP4 version ^[1]. The model is available on Tinker for fine-tuning and in the Inkling Playground, and through Together AI, Fireworks, Modal, Databricks and Baseten; it is supported by the vLLM, SGLang and llama.cpp serving stacks ^[1]. Context length options on Tinker are 64K and 256K tokens ^[1].

Place in the line

Inkling is the starting point of Thinking Machines Lab's model family rather than its end point: a smaller sibling, Inkling-Small, followed on 31 July 2026 with 276 billion total and 12 billion active parameters ^[1][2]. The release is notable for combining a very large MoE design, native multimodal input and open weights under a permissive licence, together with published safety testing by outside groups.