Inkling

weights

open(Apache-2.0 on HF, BF16 + NVFP4)

data

closed(public+third-party+synthetic, not redistributed)

code

partial(day-0 transformers/SGLang/vLLM/llama.cpp

license

Apache-2.0(OSI

Inkling's weights are released under Apache-2.0, but the training data and the full training pipeline are withheld, so it is an open-weights release rather than a fully open model. Downloading the weights also binds the user to Thinking Machines' Model Acceptable Use Policy, which restricts conduct such as weapons development and surveillance but sets no user, revenue or commercial ceiling, and which the company may revise. Press coverage often calls Inkling open source because of the license tag, but the data and training code needed to reproduce it are not public.

• https://huggingface.co/api/models/thinkingmachines/Inkling recorded 2026-09-29

  Hub API record lists weight shards such as "model-00001-of-00108.safetensors" (108 in all), carries the license:apache-2.0 tag, and reports gated as false.

• https://huggingface.co/thinkingmachines/Inkling recorded 2026-09-29

  Hugging Face model page for the BF16 release listing "975B total, 41B active" and stating that "Training data for the model was drawn from publicly available sources, acquired from third-parties, or synthetically generated or augmented"; no dataset is linked. Searched the body for gated, agree to share and access request - no access form.

• https://huggingface.co/thinkingmachines/Inkling/raw/main/README.md recorded 2026-09-29

  Raw model card front matter reads "license_link: https://www.apache.org/licenses/LICENSE-2.0" and the card lists local deployment through "Inkling supports local deployment using the following open-source libraries" (SGLang, vLLM, TokenSpeed, Unsloth, Hugging Face), inference recipes only.

• https://thinkingmachines.ai/model-acceptable-use-policy/ recorded 2026-10-11

  Model Acceptable Use Policy stating "By accessing, downloading, or using any Model Materials, you agree to be bound by this Model AUP" with conduct restrictions (illegal activity, weapons, CSAM, surveillance); no user-count, revenue or commercial-use ceiling. Dated "Last updated July 15, 2026", and "Your continued use of the Model Materials after any update constitutes acceptance of the revised Model AUP".

• https://thinkingmachines.ai/model-card/inkling/ recorded 2026-09-29

  Vendor model card gives the license as Apache 2.0 in its general-information table, says "The weights are available for download through" Hugging Face, and describes the corpus only as "Training data for the model was drawn from publicly available sources".

• https://thinkingmachines.ai/news/introducing-inkling/ recorded 2026-09-29

  Launch post stating "releasing a model we trained from scratch with the full weights available" and naming day-0 integration partners - the vendor helped "provide open-source inference and RL support in" SGLang and Miles, plus vLLM, TokenSpeed, llama.cpp (via Unsloth) and transformers - with no pretraining code release.

• https://thinkingmachines.ai/training-data-documentation/ recorded 2026-09-29

  Vendor training-data page describing sources in general terms - "This information reflects general practices, not specific to any single model" - with no corpus, listing or sampling released.

Adoption is measured as Hugging Face downloads summed across the shipped checkpoints, the BF16 release and its NVFP4 quantization, so the family total is used rather than the headline repository alone.

• https://huggingface.co/api/models/thinkingmachines/Inkling recorded 2026-08-13

  96,344 downloads in the trailing 30 days for thinkingmachines/Inkling

• https://huggingface.co/api/models/thinkingmachines/Inkling-NVFP4 recorded 2026-08-13

  272,565 downloads in the trailing 30 days for thinkingmachines/Inkling-NVFP4

On its own evaluation table Inkling beats Nemotron 3 Ultra on SWE-bench Verified (77.6 vs 70.7) but trails Kimi K2.6 (80.2) and DeepSeek-V4-Pro (80.6), and Moonshot's newer Kimi K3 scores within a few points of leading proprietary models. Thinking Machines' launch post says Inkling is not the strongest overall model available, open or closed.

• https://artificialanalysis.ai/models/inkling recorded 2026-09-29

  Artificial Analysis page on "Artificial Analysis Intelligence Index v4.3.2": "Inkling (xhigh) scores 25 on the Artificial Analysis Intelligence Index, placing it well above average among comparable models (median: 18)".

• https://huggingface.co/moonshotai/Kimi-K3/raw/main/README.md recorded 2026-09-29

  Peer check - Moonshot Kimi K3 card, "Kimi K3 is an open-weight, native multimodal agentic model", with an evaluation table set against Claude Fable 5 and GPT-5.6 Sol; notes say "Kimi K3 is evaluated with the Kimi Code harness".

• https://huggingface.co/thinkingmachines/Inkling recorded 2026-09-29

  Hugging Face model page carrying the vendor evaluation table, "Inkling results are reported at effort=0.99", with Inkling at SWE-bench Verified 77.6, SWE-bench Pro 54.3, GPQA Diamond 87.2 and AIME 2026 97.1 against Nemotron 3 Ultra, Kimi K2.5, Kimi K2.6, GLM 5.2, DeepSeek V4 Pro and closed peers.

• https://thinkingmachines.ai/model-card/inkling/ recorded 2026-09-29

  Vendor model card evaluation table (Inkling at effort=0.99) with a footnoted row, "Read the SWEBench Verified evaluation note", listing SWEBench Verified 77.6% for Inkling beside 70.7% for Nemotron 3 Ultra, 80.2% for Kimi K2.6 and 80.6% for DeepSeek V4 Pro.

• https://thinkingmachines.ai/news/introducing-inkling/ recorded 2026-09-29

  Launch post stating "Inkling is not the strongest overall model available today, open or closed" and that it "scores well among open-weights models on most agentic benchmarks".