Inkling 2 Mini: a 20GB open model NZ firms can run in-house

By DataForge team23 November 20265 min read

Thinking Machines' Inkling 2 Mini is an Apache 2.0 open-weights model that runs on a 32GB Mac, putting capable private AI on hardware NZ firms already buy.

What Thinking Machines released

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini, an open-weights model designed to run on a single machine. The Hugging Face model page lists 26.4 billion total parameters, with 3.4 billion active per token. The checkpoint is a single file of about 20GB, and the weights are published under the Apache 2.0 licence. The supported hardware includes one NVIDIA DGX Spark or an Apple silicon Mac with 32GB or more of memory. The page recommends 32GB of VRAM for other GPUs.

The model uses a looped design. Its middle layers run three times for each token, which brings the effective active parameter count to about 5.4 billion. VentureBeat's coverage reports a 1 million token context window and support for image input. It also notes that Ollama, llama.cpp and vLLM can run the model from day one. Thinking Machines describes the model as "small enough to run on your laptop, powerful enough for agentic work."

Artificial Analysis gives Inkling 2 Mini a score of 20 on its Intelligence Index. The score matches Qwen3.8 27B, which needs roughly five times as many active parameters. Artificial Analysis found that Inkling 2 Mini produced about 38% more output tokens per task than Qwen3.8 27B. The independent tester still measured the new model at roughly four times faster per task (about 9 minutes against 38) and about six times cheaper.

Why the release matters for data sovereignty

Most NZ businesses that use AI today send prompts and documents to a model hosted offshore. Every one of those requests becomes a disclosure question under the Privacy Act. Information privacy principle 12 sets conditions for sending personal information to an overseas person or entity. One condition is a reasonable belief that the recipient protects the information to a comparable standard. A contract or a hosting region can support that belief. The contract and the region still leave the data in someone else's hands.

A model that runs on a machine in your own office or rack keeps prompts, documents and outputs inside your network. The offshore disclosure question then mostly goes away for that workload. Self-hosting has been possible for a while. Until recently, though, models good enough for real work needed server-class GPUs costing tens of thousands of dollars, or they came with licences that legal teams queried. Inkling 2 Mini changes both factors at once. The hardware is a high-memory Mac or a single workstation GPU, and the licence is a standard permissive one that procurement teams already know.

The release adds to a run of capable small open models, including the Qwen3.8 27B model Artificial Analysis used for comparison. For an NZ firm, the practical effect is that a private deployment no longer forces a big drop in quality, at least for the tasks this model handles well.

Where Inkling 2 Mini is weak

Thinking Machines is unusually candid about the model's gaps, and NZ buyers should read the model card limitations before piloting anything. The card says the model knows less than its larger sibling, Inkling 2 Small, and may state wrong facts confidently. The card also warns that the model is less reliable with tools and may take destructive actions, such as deleting files, without asking for confirmation. Long reasoning loops are a known problem, so the company recommends caps on output tokens and agent turns.

The launch post says looping improves reasoning without adding stored facts, and that Qwen3.8 27B beats Inkling 2 Mini on multi-turn chat and complex instruction following. Inkling 2 Mini is strongest on agentic coding, search and tool use. Quality drops outside English, and the model card's non-English example comes from testing in Chinese. Thinking Machines hasn't published any results for te reo Māori, so organisations that handle te reo content should test that use case themselves.

The licence also carries conditions. Apache 2.0 covers the weights, while a separate Model Acceptable Use Policy applies to anyone who downloads or runs the model, including on their own hardware. The policy restricts consequential decisions and unlicensed professional advice without human oversight. A business that offers the model to its customers becomes responsible for those customers' compliance and has to disclose known limitations to them. Thinking Machines also runs a hosted API through its Tinker service. The launch materials don't describe how Tinker retains or handles customer data, so the hosted option needs its own privacy review before any personal information goes near it.

What we'd expect for NZ businesses

Our view is that Inkling 2 Mini suits contained internal workloads where confidentiality matters more than broad general knowledge. Good candidates include summarising internal documents, drafting from your own templates, searching a document store through retrieval, and helping developers on code that can't leave the building. Each of those tasks gives the model the facts it needs, which reduces its weak factual recall.

We'd keep the model away from unsupervised agent work on production systems for now. Its own developer flags tool reliability and destructive actions as known weaknesses. Any agent built on Inkling 2 Mini needs narrow permissions, confirmation steps for anything that writes or deletes, and logs that a person reviews.

Running a model in-house also moves work onto your team. Someone has to patch the host machine, control who can reach the model, keep logs, and update the weights when Thinking Machines ships fixes. A self-hosted model with weak access controls can leak data inside the business just as easily as a cloud service can leak it outside. We cover the hardware and operating side in our local AI server work. Choosing which use cases justify that effort is part of AI adoption and governance planning.

Next steps

  • List the AI tasks you currently send offshore, and mark the ones that involve personal, commercial or client-confidential information.
  • Pick one marked task that runs on internal documents, and trial Inkling 2 Mini on a 32GB Mac or a single GPU workstation with test data first.
  • Score the trial outputs against your current cloud tool using real examples, and check factual accuracy and any te reo Māori content by hand.
  • Have someone read the Acceptable Use Policy before deployment, especially if customers or the public will interact with the model.
  • Set output token and agent turn limits, and require a person to confirm any action that writes, sends or deletes.
  • Record where the weights came from, which version you run, and who can access the machine, then add the model to your IT asset and change processes.
  • Keep the hosted Tinker API out of scope for personal information until Thinking Machines publishes clear data handling terms.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS