Thinking Machines Lab has released Inkling 2 Mini, an open-weight AI model designed for tasks including coding, tool use and work with images and text. The company says the model can run locally on supported hardware, offering developers another way to experiment without sending every request to a hosted service.
The November 20, 2026 release is notable for its design as well as its availability. Inkling 2 Mini repeats part of its neural network during each generation step, a technique called looping. It also uses a mixture-of-experts architecture, selecting a subset of its components as it processes information. The result is a model that stores more parameters than it activates in a single pass, although looping still adds computation.
There is a practical catch behind the word mini: the official model card describes a roughly 20 GB checkpoint and recommends substantial memory. Local access avoids a per-request hosted API bill, but hardware, electricity, setup and performance constraints still matter. This is an option for capable local machines, not a promise that any phone or laptop can run a million-token conversation.

What Thinking Machines released
Inkling 2 Mini accepts text and image inputs and generates text. The developer's announcement and model card describe it as a general-purpose model for applications such as coding assistants, chatbots, retrieval-assisted systems and agents that use tools. The weights are available on Hugging Face under the Apache 2.0 license, with an acceptable-use policy also linked from the model card.
Thinking Machines also offers access through Tinker for inference and customization, and identifies a playground for chat. Local self-hosting and a hosted platform are different arrangements. The availability of downloadable weights does not establish that every hosted request is free, and readers should check current service terms and pricing before using an API.
Hugging Face published a technical walkthrough on November 20 with support details for Transformers, llama.cpp, MLX and SGLang. Thinking Machines also lists several other inference tools, including Ollama on Apple silicon and quantized GGUF options. Because tool support evolves, check the current documentation for the exact operating system and model format before installing anything.
The release attracted independent technology coverage on November 21, including The Next Web's report. Its significance for readers is straightforward: an open-weight model lets a developer examine available files and choose where to run them, subject to the license and practical hardware limits.
Why the model loops
Think of a language model as a collection of layers that transform information before it generates the next piece of text. Inkling 2 Mini has 24 stored decoder layers, according to the model card. Its middle 12 layers are run three times using shared weights, producing 48 layer applications per token in the standard configuration.
That arithmetic is six initial layers, plus twelve middle layers visited three times, plus six final layers: 6 + 36 + 6 = 48. It does not mean the model contains 48 separately stored layers. Reusing weights can change the balance between how much model data must be stored and how much work is performed while generating an answer.
Looping is not a free speed boost. Each additional pass performs work, and actual performance depends on the implementation, hardware, prompt length and other settings. Hugging Face notes that each layer visit also maintains its own attention and related state. Someone evaluating a model should measure the full workload rather than assuming a smaller checkpoint automatically makes every request faster.
The lab also released a research-specific variable-loop model. Hugging Face says it can be run with different numbers of passes, but describes its own testing as limited. That variant is useful for studying trade-offs; it should not be confused with a guarantee that changing loop count will preserve output quality in a production application.
What mixture of experts means
Thinking Machines lists 26.4 billion total parameters and 3.4 billion active parameters for Inkling 2 Mini. Its model card explains that accounting for repeated loop passes gives about 5.4 billion effective weight activations per token. In each layer, the routing mechanism selects eight of 128 experts for a token.
These figures describe different things. Total parameters help characterize the full model being stored. Active parameters describe a subset used during a pass, while effective activations account for repeated work. None of those numbers alone tells a buyer exactly how much memory a running application will use or how accurate it will be.
- Inputs and output: text and images in; text out.
- Architecture: 24 stored layers, with the middle 12 visited three times in the standard model.
- Parameters: 26.4 billion total and 3.4 billion active, with about 5.4 billion effective activations when looping is included.
- Context: up to one million tokens is listed; achievable settings depend on the runtime and hardware.
- Weights: downloadable under Apache 2.0, with the linked use policy worth reviewing.
Can it run on an ordinary computer?
The answer depends on the machine. The official card lists one NVIDIA DGX Spark, Apple silicon devices with at least 32 GB of memory, and other GPU platforms for which it recommends 32 GB or more of video memory. The company's launch announcement specifically describes Ollama support on Macs meeting its memory requirement.
A checkpoint of about 20 GB is only the starting point. The operating system, inference software, input data and memory needed to process a conversation also consume resources. A one-million-token advertised context is a model capability, not evidence that a particular 32 GB computer can comfortably process a full million-token prompt at a useful speed.
Quantized versions can reduce the space needed for model weights by storing numbers at lower precision. Hugging Face describes four-bit and eight-bit GGUF versions for compatible tools. Quantization can change memory needs and output quality, and the exact trade-off depends on the version and runtime. Before using an unofficial conversion, verify its publisher and read its documentation.
For a Windows desktop with less dedicated GPU memory, partial CPU use or alternative quantization may be technically possible in some runtimes, but the published support list does not guarantee a particular speed or successful configuration. The right first step is to inspect the current runtime's requirements and the model card, not to purchase hardware based on a headline.
What about coding performance?
Thinking Machines emphasizes agentic coding, in which an AI system may inspect a repository, propose edits, run tools and respond to feedback. In its own reported evaluations, the lab lists a 48.2% result on DeepSWE v1.1 and a 79.2% result on SWE-bench Verified using the specified setup. These are developer-reported scores for named tests, not guarantees about a reader's own software project.
Benchmarks can be helpful when their settings are clear, but comparison is complicated by differences in prompts, tools, test versions and evaluation harnesses. The company itself notes that evaluation setups vary across models. A score obtained with one agent workflow should not be presented as a universal ranking for every coding assistant.
Hugging Face says its informal tests found the model at least comparable to Qwen 3.8 27B, while explicitly calling for more validation. The Next Web also reported an Artificial Analysis index result. These are useful early signals, but developers should still test their own languages, repositories and constraints before moving production work to a new model.
Practical evaluation can begin with tasks that already have known answers: explain a small module, write tests for a contained function, or propose a patch in a disposable branch. Check the diff, run the existing tests and inspect whether the model follows project instructions. An agent that produces convincing code can still introduce a bug, leak a secret or misunderstand a requirement.
The limitations the developer disclosed
Thinking Machines says Inkling 2 Mini is not its strongest small-model option for multi-turn chat or instructions containing many constraints when compared with some models in its published tables. The announcement also says it trails larger models on certain reasoning-heavy tests. Those qualifications matter when a model is being promoted for autonomous tasks that may require consistent behavior across many steps.
The developer's model card recommends human oversight for high-stakes use, testing for a specific application, and sandboxing agentic workloads. It also recommends requiring confirmation before destructive or irreversible actions. Downloading weights gives users more control over deployment; it does not remove the responsibility to secure the surrounding application.
Looping also raises a monitoring question. The lab says additional internal computation can occur without appearing as visible reasoning tokens. It has published research on the effect of loop count on monitorability using a research variant. Those are the lab's findings under its experimental conditions, not proof that all unsafe behavior is observable or that every deployment is secure.
How to try it without promising free cloud access
A cautious reader has two different paths. The first is to use the downloadable model with a supported local runtime. That may avoid a hosted inference charge for requests executed on the user's own machine, but the download, computer, energy and maintenance still have costs. The second is a hosted service such as Tinker, where availability, limits and pricing depend on the provider's current terms.
For local testing, start with the official model card and the Hugging Face inference guide. Select a supported runtime, check its hardware and software requirements, and obtain model files from the named official or trusted publishers. Keep the model server bound to the local machine unless you have deliberately secured network access.
Then run a small, non-sensitive prompt. Record startup time, memory use and output speed before increasing context length or adding image inputs. If you connect the model to a coding agent, use a sandbox or a disposable workspace first and review the agent's permissions. A local model is not automatically safe to grant access to passwords, private files or unrestricted command execution.
For readers looking specifically for free API credits, the launch materials reviewed here do not establish an unlimited free hosted API. Check the live provider pages and account terms. Self-hosting can be a useful alternative to token billing, but it is not the same thing as unlimited free computing.
Why the release matters
Analysis: Inkling 2 Mini gives developers another way to compare three decisions that are often bundled together: whether model weights are accessible, where inference runs, and how an application routes tasks. Open weights can support local experimentation and customization; the looped design offers a different compute trade-off; hosted access offers convenience with its own costs and data-handling terms.
The strongest evidence in the announcement concerns what was released and how the lab says it was built. The broader question is how the model performs on real work outside the developer's examples. Independent measurements of memory, latency, code quality and reliability across different languages and hardware will help answer that.
For now, the most useful conclusion is neither that a small model has made larger systems obsolete nor that local AI is costless. A new downloadable option has arrived, with a published design, documented hardware guidance and identifiable limits. Developers can test it against the work they actually need to do.
Sources and reporting notes
This report uses Thinking Machines Lab's November 20 announcement and model card, Hugging Face's technical guide and independent coverage by The Next Web. Benchmark results are attributed to their publishers. Availability, prices and software support may change; the feature image is an editorial illustration.
- Thinking Machines Lab: Introducing Inkling 2 Mini, November 20, 2026.
- Thinking Machines Lab: Inkling 2 Mini model card.
- Hugging Face: Inkling 2 Mini technical guide, November 20, 2026.
- The Next Web: Inkling 2 Mini release coverage, November 21, 2026.


