TECH Nov 22, 2026 · by Skills Team

Inkling 2 Mini: Small AI Model That Codes Locally

Inkling 2 Mini is a 26.4 billion parameter open weight model from Thinking Machines Lab that runs on a 32GB Mac or a single DGX Spark and scores 79.2 percent on SWE-bench Verified. This guide explains how its looped design works and how teams can evaluate it.

Inkling 2 Mini: Small AI Model That Codes Locally

Key Takeaways

  • Inkling 2 Mini is a 26.4 billion parameter open weight model that runs on a 32GB Mac or a single DGX Spark.
  • Thinking Machines Lab reports a 79.2 percent score on SWE-bench Verified for agentic coding tasks.
  • A looped transformer design reuses middle layers three times, so a small model reasons longer per token.
  • Teams should pilot the model on real repositories and measure review effort before wider adoption.

Inkling 2 Mini is a small open weight AI model from Thinking Machines Lab that codes at a level previously seen only in far larger models, and it runs on hardware a team can own outright. Released on November 20, 2026, the model packs 26.4 billion total parameters into a checkpoint of about 20GB, activates only 3.4 billion parameters per token, and scores 79.2 percent on SWE-bench Verified, a respected benchmark for real world software repair tasks. For business teams, the significance is simple: capable coding assistance no longer has to live on someone else's servers. This article explains how the looped transformer design works, what the benchmark results really mean, where local models fit into coding workflows, and how a team can evaluate Inkling 2 Mini before rolling it out.

What Is Inkling 2 Mini?

Inkling 2 Mini is a general purpose multimodal AI model from Thinking Machines Lab that accepts text and image inputs and produces text outputs. It was released on November 20, 2026, as an open weight model under the Apache 2.0 license, with downloadable weights on Hugging Face and hosted inference and fine tuning through the lab's Tinker platform. The model is built for agentic coding, tool use, and document work, the kind of multistep tasks where a model plans, acts, checks results, and adjusts. Its headline numbers are striking for its size: 26.4 billion total parameters, only 3.4 billion active per token, a checkpoint of about 20GB, and a context window of up to 1 million tokens when run locally.

According to the Hugging Face model card for Inkling 2 Mini, the model ships as a single checkpoint of about 20GB and supports local inference on one NVIDIA DGX Spark or on Apple silicon devices with 32GB or more of memory. The gap between total and active parameters is the point. A mixture of experts design stores many specialist sub-networks, called experts, and activates only a small subset for each token, so most of the stored knowledge stays idle while the model works. Thinking Machines Lab also reports about 5.4 billion weight activations per token once repeated passes are counted, still a small fraction of the stored model.

The model arrived shortly after its much larger sibling, Inkling 2 Small, a 594 billion parameter model, and it takes the opposite approach: instead of adding parameters, it reuses them. Day one runtime support is broad, including Ollama, vLLM, SGLang, llama.cpp, Unsloth, NVIDIA Dynamo, TokenSpeed, and the Llama App, so teams can try it with familiar tools. That combination of small size and wide tool support is what makes the model interesting to teams that want to run AI on their own machines.

How Does Inkling 2 Mini Work?

The model works by running the same middle layers several times for every token instead of passing each token through every layer once. The model stores 24 transformer layers arranged in four blocks of six, and the two middle blocks run three times per token, producing 48 layer applications from 24 stored layers. In plain terms, the model spends more internal effort on reasoning before it produces its next word, without storing extra weights to do it.

A looped transformer is a transformer architecture in which some layers are reused in repeated passes over the same token, trading extra computation for a smaller stored model. That trade suits coding well, because coding rewards careful reasoning about a problem more than it rewards memorised facts. The repeated passes give the model more chances to refine its understanding of the code it is reading before it writes anything, which helps with tasks that need several steps of logic rather than a single quick answer.

The checkpoint stays small for a second reason: mixed precision. Twenty layers are stored natively in NVFP4 precision and the last four in MXFP8, so the weights are compact from training rather than squeezed down afterwards. The result is a model that fits in about 20GB while keeping most of its accuracy. One caveat comes from the model card itself, which warns that support for looped models in local runtimes is new and speed varies by runtime and hardware, so early users should test the runtime they plan to use.

Why Does Inkling 2 Mini Matter for Business Teams?

This model matters for business teams because it moves capable coding assistance onto hardware the business controls, which changes who holds the code, the prompts, and the working context. When a model runs locally, source code and internal documents do not need to leave the building to get AI help, and work can continue without a network connection. Thinking Machines Lab demonstrated this by running the model offline on a MacBook Pro during a flight, where it built a playable flight simulator in a few minutes.

Thinking Machines Lab reports that Inkling 2 Mini scores 79.2 percent on SWE-bench Verified and 77.9 percent on SWE-bench Pro v2 Hard, with both figures based on the lab's own evaluations. SWE-bench Verified measures a model's ability to fix real issues from open source repositories, so a strong score there is a useful signal for coding work, even though it is not a guarantee of results on a private codebase. For teams, the practical read is that a small local model can now handle a meaningful share of routine coding assistance.

Local models also fit naturally into a wider self-hosted AI strategy, where a business keeps sensitive work in house and uses hosted services only where they add clear value. Our guide to self-hosted AI models for business explains how teams weigh that choice, and an IT consulting engagement can map which workloads belong on local hardware and which do not. The point is not to replace every hosted tool at once, but to gain a real option for work that should stay under the team's own control.

What Can Inkling 2 Mini Do for Coding Workflows?

The model can handle the multistep coding work that teams actually hand to AI assistants: exploring a repository, planning a fix, editing files, running tests, and revising until the tests pass. Its strongest results are in agentic coding, where the model uses tools, reads files, and iterates, rather than in single turn question answering. A context window of up to 1 million tokens when run locally also means it can hold large parts of a codebase in view at once, which matters for changes that touch many files.

Futurum Group, reporting the lab's figures, notes that Inkling 2 Mini scores 79.0 percent on BrowseComp, 85.3 percent on DeepSearchQA, and 82.7 percent on MCP Atlas, benchmarks that test agentic search and tool use. Those results point to a model that can gather information, call tools, and work through a task in stages, which is the shape of most real development work. A developer might ask it to track down a bug across several files, draft the fix, and check the result against the test suite.

Everyday uses go beyond bug fixing. The model can draft boilerplate code, write unit tests, explain unfamiliar code, produce documentation, and review changes for common mistakes. Testing remains essential, because AI generated code still needs human review and automated checks before it ships. Our QA playbook for testing AI generated code sets out the checks that catch the mistakes models make most often, from edge cases to silent failures in logic.

How to Evaluate Inkling 2 Mini for Your Team

A team should evaluate the model on its own real work, not on benchmark tables alone, because benchmarks measure public repositories while a business cares about its private codebase. A short structured pilot gives a clear answer in days rather than weeks. Start with a real repository and a small set of tasks the team already understands, then compare the output against the team's normal standard for code quality.

  1. Pick one real repository and five to ten recent tasks, such as bug fixes or small features, and run them through the model.
  2. Record how much review and rework each output needs, and how often the first attempt passes the existing tests.
  3. Check the hardware the team already has, since local runs need a 32GB Mac, a DGX Spark, or a GPU with 32GB or more of memory.
  4. Confirm the runtime, such as Ollama or vLLM, and note the context limit in use, since hosted runs through Tinker cap context at 256K tokens while local runs can reach 1 million.
  5. Review the Apache 2.0 license terms and the team's data handling rules before any internal code is used with the model.

Many AI pilots stall because they are tested on toy problems instead of real work, and the pilot never turns into a production tool. Our guide on why AI pilots never reach production explains the traps that keep pilots from becoming production tools, including unclear ownership and no agreed measure of success. A pilot should therefore end with a written decision: adopt the model for a defined set of tasks, keep testing, or stop. Without that decision, even a good pilot fades away.

What Are the Limits of Inkling 2 Mini?

The model has real limits, and the honest ones start with the evidence itself: the headline benchmark scores are reported by Thinking Machines Lab, and independent reproduction is still catching up. Early independent coverage also notes that the model trails Qwen3.8-27B on multiturn chat and constraint heavy instructions, so teams that need long polished conversations or strict instruction following should test those cases directly rather than assuming coding strength carries over.

Local runtime support for looped models is new, so speed and stability can vary between runtimes and hardware, and the 20GB checkpoint still needs a machine with at least 32GB of memory. Hosted runs through Tinker cap context at 256K tokens, well below the 1 million token ceiling available locally. Small models can also produce confident but wrong code, so human review and automated tests stay mandatory. The model is best treated as a capable assistant for defined coding tasks, not as a replacement for senior judgement.

Frequently Asked Questions

What hardware do I need to run Inkling 2 Mini locally?

Inkling 2 Mini needs a machine with at least 32GB of memory, according to the official model card. Supported setups include one NVIDIA DGX Spark, Apple silicon Macs with 32GB or more of memory, and other GPUs with 32GB or more of VRAM. The checkpoint is about 20GB, so the extra memory covers working space and the context window.

Can Inkling 2 Mini really code as well as larger models?

On coding benchmarks, Inkling 2 Mini performs close to larger models, with Thinking Machines Lab reporting 79.2 percent on SWE-bench Verified. Those are the lab's own results, and independent checks are still in progress. Teams should pilot the model on their own repositories before assuming it matches larger models on every task.

What license does Inkling 2 Mini use?

Inkling 2 Mini is released under the Apache 2.0 license, with weights available on Hugging Face. The Apache 2.0 license is a permissive open source license that allows commercial use, modification, and distribution under its terms. Teams should still have their own legal review confirm the terms fit their plans before using the model in commercial products.

Can Inkling 2 Mini work offline?

Yes, Inkling 2 Mini can run fully offline on local hardware, since the weights run on the team's own machine. Thinking Machines Lab demonstrated this by running the model on a MacBook Pro in airplane mode during a flight. Offline use keeps code and documents on local hardware, which suits teams with strict handling rules for internal work.

How should a business start using Inkling 2 Mini?

A business should start with a small pilot on one real repository, using tasks the team already understands and can judge. Measure review effort and test pass rates, then decide whether to adopt it for defined tasks. Keep human review and automated tests in place throughout the pilot, and set a clear decision date before the pilot begins.

Small capable models that run on a team's own hardware change the shape of coding assistance, and this release is a strong early example of that shift. Skills Technology Solutions, a software agency based in Rahim Yar Khan, Pakistan, serving clients in Pakistan and the UK, can help you put this into practice. Contact our team to discuss your project.

#Inkling 2 Mini#AI coding#open weight AI#local AI models#Thinking Machines Lab