By Mahan Kenway
Disclosure: This article was researched and written by an autonomous AI system for Mahan Kenway. Cover: original AI-generated conceptual illustration, not a product photograph or architecture diagram.
A model can fit on a laptop and still ask that laptop to do a lot of work. That distinction is central to Inkling 2 Mini, an open-weights multimodal model released by Thinking Machines Lab on November 20.
The company says its new model can run locally, accepts text and images, and is aimed at tasks such as coding and tool use. Its most interesting design choice is a loop: the model revisits a set of internal layers three times for each token, reusing weights instead of storing a separate layer for every pass. That can put more computation behind each output without multiplying the stored weights by the same amount.
For developers who prefer local inference, the release is worth a look. But “runs on a laptop” does not mean “runs well on every laptop,” and benchmark scores from agent setups should not be mistaken for a guarantee that an untested coding assistant will safely edit a production repository.
The numbers behind the loop
Inkling 2 Mini has 26.4 billion total parameters, according to its model card. It is a mixture-of-experts model: rather than using every expert for every token, it routes each token to a subset. The card lists 3.4 billion active parameters and roughly 5.4 billion effective active parameters per token after accounting for the extra loop passes.
There are 24 physical decoder layers. The first six run once, the middle 12 run three times, and the final six run once: 48 layer applications using the weights of 24 layers. The Hugging Face technical walkthrough explains how later passes receive messages derived from earlier ones. A separate research checkpoint allows a variable number of passes; the main Mini release uses three.
This is not a free performance trick. Reusing weights saves storage compared with adding entirely separate layers, but those passes still consume computation and memory bandwidth. The model is doing more work per generated token, not magically turning 26.4 billion parameters into a larger model at no cost.
How local is “local”?
The official checkpoint is about 20 GB in a mixed NVFP4 and MXFP8 format. Thinking Machines lists Apple silicon computers with at least 32 GB of unified memory and an NVIDIA DGX Spark among its supported local hardware, and recommends 32 GB or more of VRAM for other GPUs and platforms. Its demonstration used a MacBook Pro with 48 GB of memory in airplane mode. These are examples and recommendations, not promises of a particular speed on your hardware.
The model accepts text and images and produces text; its advertised context window is up to one million tokens. That upper limit is not a promise that a million-token session will fit comfortably on a 32 GB machine. Runtime state, the context cache, the serving framework, other applications, and the chosen quantization all affect memory use and speed. Start with a modest context and measure your own workload.
The weights are released under Apache 2.0, according to the model card. Downloading a model to your machine can keep inference local, but privacy depends on the whole setup: a coding agent with web search, remote APIs, telemetry, cloud storage, or unrestricted shell access can still send or expose data outside the machine. Inspect the tools and permissions, not just where the weights live.
What the benchmarks suggest, and what they don’t
Thinking Machines reports 48.2% on DeepSWE v1.1 using mini-swe-agent and 79.2% on SWE-bench Verified with a bash-only setup. Those figures describe specific tests and configurations, not the percentage of your own bugs it will fix. Agent harnesses, tool access, prompts, inference settings and the selected evaluation set all matter.
There is some independent context. Artificial Analysis reports that Inkling 2 Mini scored 20 on its Intelligence Index, matching its measured Qwen3.8 27B and Nanbeige4.3-33B results. Its compute comparison uses an active-parameters-times-tokens proxy, not a direct measurement of your laptop’s energy use or latency. The evaluator also reports that Inkling used more output tokens than Qwen on its index, even as its proxy estimated less compute. That is a good reason to compare end-to-end task cost and latency rather than assuming fewer active parameters always means faster answers.
The model card itself shows uneven results: Inkling 2 Mini ties Qwen3.8-27B on Terminal Bench v4.0 in the listed setup but trails it on Terminal Bench v2.1. Thinking Machines also notes that looping does not simply add factual knowledge. Treat plausible explanations and generated code as hypotheses to check, particularly when the task depends on facts unavailable in the prompt.
A practical first try
On a supported Apple silicon machine with an up-to-date Ollama installation, the model registry gives this command:
ollama run inkling-2-mini The first run needs a network connection to download the model. For a Linux or command-line server route, Hugging Face documents a 4-bit GGUF build with a compatible up-to-date llama.cpp installation:
llama serve -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL These are different runtimes and potentially different model formats, so check their documentation for current hardware and feature support. Do not assume an image-input workflow or a token-per-second result transfers unchanged between them.
A sensible evaluation is a small set of tasks you already understand: explain a function, write tests for an existing bug, and propose a patch in a disposable checkout. Record the runtime, memory consumption, and whether the tests pass; inspect the diff yourself. If you connect an agent, limit filesystem and network permissions and avoid giving it secrets by default.
Inkling 2 Mini makes a meaningful tradeoff: reuse parameters to spend more computation on each token while keeping an open-weight model small enough for capable personal hardware. Its real appeal is the option to test that tradeoff on your own tasks, under your own privacy and resource constraints. The result to trust is the one you can reproduce in your environment, not a headline benchmark alone.
