Thinking Machines releases Inkling 2 Small (594B MoE, 17B active) with Apache 2.0 open weights due next week

Overview

Thinking Machines Lab, the AI company founded by former OpenAI Chief Technology Officer Mira Murati, released Inkling 2 Small on 5 November 2026. It is the first model in the company’s second-generation family. According to VentureBeat, Inkling 2 Small is a sparse Mixture-of-Experts (MoE) model with 594 billion total parameters, 17 billion of which are activated for each generated token. It accepts text and image inputs and supports a context window of up to 1 million tokens. The company has said it will publish the model weights under an Apache 2.0 licence the following week, once external red-teaming and safety evaluations are complete.

For Australian professional services firms, the licence matters more than the benchmark scores. Until now, near-frontier AI performance has generally meant sending documents to a proprietary model through a vendor’s API. An open-weight model under a permissive licence can be run on hardware the firm controls, either on premises or in a private cloud environment hosted in an Australian region. For environmental consultancies, law firms, developers and councils that handle confidential transaction material, privileged legal advice and personal information, that changes where data sits and who can access it.

The reported results need to be read in context. Inkling 2 Small was evaluated on software engineering and enterprise agent tasks such as procurement, billing and operational troubleshooting. None of the published benchmarks test environmental, scientific or regulatory judgement. This article sets out what was released, what the reported numbers mean, and what Australian firms should weigh before treating a self-hosted model as a solution to their data handling concerns.

Key details

Inkling 2 Small uses a sparse Mixture-of-Experts architecture with 594 billion total parameters and 17 billion active parameters per token. In an MoE design, only a subset of the model’s internal “experts” process each token, so the computation per token resembles that of a much smaller dense model. The full set of 594 billion parameters still has to be held in memory, however. The model handles text and high-resolution images but has no native audio input. Users can select the reasoning effort, trading response speed against more exhaustive checking of the model’s own work. Training reportedly drew on codebases, terminal execution environments, visual document analysis and multi-turn agent interactions, and used compute clusters that included NVIDIA GB300 NVL72 hardware.

Independent evaluator Artificial Analysis gave Inkling 2 Small a score of 31 on its Intelligence Index, up from 13 for the previous Inkling Small. According to VentureBeat, once the weights are published, that score would make it the highest-ranked open-weight foundation model developed outside China on the index at the time of reporting, ahead of Mistral Large 4 (nicknamed “Le Chonk”), which scored 26. On task-specific evaluations, the model reportedly scored:

  • 71.8% on Surge AI’s EnterpriseBench-CoreCraft, a third-party test of autonomous agents working through enterprise procurement, billing and operational troubleshooting tasks
  • 86.3% on SWEBench Pro v2 (Hard), a software engineering benchmark
  • 72.1% on DeepSWE v1.1, another software engineering agent evaluation

Thinking Machines also reported a fine-tuning result. In evaluation runs on financial and corporate reporting documents, Inkling 2 Small matched the final performance of the previous-generation model after only 75 training steps, then continued to improve. Fast convergence during fine-tuning reduces the compute time and engineering effort needed to adapt a model to an organisation’s own document formats, templates and data schemas. This figure comes from the developer rather than an independent evaluator, and it relates to financial documents rather than technical or scientific reports.

At release, the model is available only through Thinking Machines’ Tinker platform, which supports inference and custom fine-tuning. The weights remain closed until the safety evaluation is finished. Despite the “Small” label, which reflects the company’s internal model tiering, self-hosting is a serious infrastructure undertaking. As a simple calculation, 594 billion parameters stored at 8-bit precision occupy roughly 594 GB, and at 16-bit precision roughly 1.2 TB, before allowing for the additional memory needed to process long inputs. Running the model in-house therefore means a multi-GPU server built on data centre accelerators such as NVIDIA Hopper or Blackwell parts, not a desktop workstation.

vibehacker.com
Image source: vibehacker.com

Australian context

For Australian firms, the main relevance of an open-weight model is control over where information is processed under the Privacy Act 1988 (Cth). Australian Privacy Principle 8 governs cross-border disclosure of personal information, and Australian Privacy Principle 11 requires entities to take reasonable steps to protect personal information from misuse, interference and unauthorised access. When a firm sends documents containing personal information to an overseas-hosted AI service, both principles come into play. A model running on infrastructure the firm controls, located in Australia, removes the cross-border disclosure step, although APP 11 security obligations continue to apply to the firm’s own systems. The Privacy and Other Legislation Amendment Act 2024 adds further transparency obligations for automated decisions that significantly affect individuals, with those provisions scheduled to commence on 10 December 2026.

Privacy law is only part of the picture for contaminated land work. Environmental due diligence reports for property transactions are often prepared under data room confidentiality terms or at the request of lawyers to support a claim of legal professional privilege. Site history material, including title searches, historical aerial photographs, council records and earlier investigation reports, can also identify current and former landowners, occupiers and other individuals. Uploading this material to an overseas AI service may conflict with confidentiality terms or create risk for a privilege claim, whether or not personal information is involved. A self-hosted model can reduce that exposure, but firms still need to check engagement terms, set access controls and keep records of how the material is used. Because none of the published benchmarks test environmental or regulatory judgement, firms should also test the model on their own representative reports, and have qualified practitioners review any AI-assisted output, before relying on it for technical work.

References and related sources

How iEnvi can help

iEnvi integrates technology and data-driven approaches into environmental consulting. We monitor AI and technology developments that affect how environmental professionals deliver services to clients.


This is an iEnvi Machete news summary. Prepared by iEnvi to summarise the source article for environmental professionals tracking AI, data, and technology developments that affect consulting and project delivery.

Published: 08 Nov 2026

Need advice on this topic? Speak to an iEnvi expert at info@ienvi.com.au or 1300 043 684, or contact us online.

Need advice on this issue? iEnvi provides practical, senior-led environmental consulting across contaminated land, remediation, ecology and environmental risk.

Contaminated land services Environmental due diligence Talk to iEnvi