Thinking Machines Lab made the weights of Inkling 2 Small downloadable on Hugging Face on November 12, 2026, a week after it first released the model through its own Tinker platform with closed weights. The model card lists 594 billion total parameters, 17 billion active per token, and an Apache-2.0 licence, which matters to research groups that need to run, inspect or fine-tune a model locally rather than call a hosted API.
What Thinking Machines announced
In its Inkling 2 Small announcement, dated November 5, 2026, the lab said the model launched on Tinker first, with weights to follow after external safety testing and red-teaming, and that it intended to release the full weights under Apache 2.0. A November 12 update on the same page states that the weights are now on Hugging Face. The page also lists availability through Unsloth, SGLang, vLLM and TokenSpeed, and hosted access through providers including NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare.
The lab describes a mixture-of-experts model that accepts image input and supports variable reasoning effort. It says the model does not support audio and is less polished for chat. It also says the model trails closed-weight models such as GPT-6.1 Sol on its own alignment evaluations and calls reducing that gap a priority for future models.
What the model card specifies
The Hugging Face model card gives the following specifications.
- Size and design: 594B total and 17B active parameters; a sparse mixture of experts that routes each token to 8 of 384 experts; 48 decoder-only layers with hybrid local and global attention.
- Inputs and outputs: text and image in, text out; context of up to 1M tokens.
- Licence: Apache-2.0.
- Numerics: native NVFP4, with an MXFP8 alternative; the card recommends NVFP4 for nearly all uses.
- Hardware: at least 500 GB of aggregate VRAM for NVFP4 (for example 2x B300 or 4x H200), and at least 800 GB for MXFP8 (for example 4x B300 or 8x H200).
- Fine-tuning: supported through Miles, SkyRL and Snowflake Cortex Training, per the card.
The organisation page on Hugging Face shows two repositories for the model, Inkling-2-Small and Inkling-2-Small-MXFP8, both listed at 594B parameters. The earlier Inkling-Small, which the page lists at 266B parameters, remains available.
What it means for researchers and research offices
This is an editorial reading of the published specifications, not a claim by the lab.
- Compute is the practical barrier. The memory figures above put local use at multi-GPU server scale. Many departments will reach the model through a hosted provider or institutional cluster rather than a single workstation, which raises data-handling questions that a local deployment avoids.
- Licence terms are permissive on their face. Apache-2.0 generally allows use, modification and redistribution, including for research. Research offices should still check the repository for any additional use policy and confirm how outputs and fine-tuned derivatives will be shared.
- Reproducibility improves with fixed weights. A downloadable checkpoint can be pinned and cited by version, unlike a hosted endpoint that may change. See CASRAI’s entries on open weights models and model cards.
- Benchmarks are vendor-reported. The announcement and card report figures such as 86.3% on SWEBench Pro v2 (Hard), 81.7% on ARC-AGI-2 and 28.7% on HLE-Diamond, and an Artificial Analysis Intelligence Index of 31 in the announcement (30.6 in the card’s table). These are the lab’s own reported results, and independent replication is the appropriate test before relying on them.
Safety and governance context
The model card says safety evaluations covered sycophancy, manipulation, refusals, CBRN, cyber and loss-of-control risks, and concludes the model does not pose material uplift beyond what the open-weight ecosystem already provides. That is the developer’s conclusion. It also lists limitations including hallucination, uneven performance across languages and domains, and degraded performance in long multi-turn conversations, and advises against medical, legal or safety-critical use without fine-tuning, validation and human oversight. For how public bodies assess releases of this kind, see CAISI’s evaluation cadence for open-weight models and the safety case for and against open-weight models. Groups choosing a model for study work can use CASRAI’s framework for choosing and governing LLMs.
Sources
- Thinking Machines Lab, Introducing Inkling 2 Small (announced November 5, 2026; updated November 12, 2026). Read 2026-11-14.
- Hugging Face, thinkingmachines/Inkling-2-Small model card. Read 2026-11-14.
- Hugging Face, thinkingmachines organisation page. Read 2026-11-14.
