{
  "id": 21282310,
  "title": "Inkling 2 Mini Brings a Looped AI Model to Local Machines. Here’s What That Actually Means",
  "url": "https://urgent.news/2026/11/22/inkling-2-mini-brings-a-looped-ai-model-to-local-machines-heres-what",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-22T20:37:42.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/mahankenway/inkling-2-mini-brings-a-looped-ai-model-to-local-machines-heres-what-that-actually-means-bf"
  },
  "original_language": "en",
  "account": "Inkling 2 Mini is a recently released open-weights multimodal model that can run locally on devices like laptops. Developed by Thinking Machines Lab, the model can process both text and images to perform tasks such as coding and tool use. Its innovative design feature is a loop, where the model passes through a series of internal layers three times for each token, reusing weights instead of creating new layers for each pass. This approach allows for increased computation without proportionally increasing the storage of weights.\n\nWith 26.4 billion total parameters, Inkling 2 Mini is a mixture-of-experts model, meaning it routes each token to a subset of experts rather than using every expert for every token. The model card indicates that there are 3.4 billion active parameters and roughly 5.4 billion effective active parameters per token, accounting for the additional loop passes. The model employs 24 physical decoder layers, with the first six running once, the middle 12 running three times, and the final six running once, resulting in 48 layer applications using the weights of 24 layers.\n\nThe official checkpoint of Inkling 2 Mini is approximately 20 GB in size, using a mixed NVFP4 and MXFP8 format. It is recommended that devices with at least 32 GB of unified memory and an NVIDIA DGX Spark, or an Apple silicon computer with 32 GB or more of VRAM, be used for optimal performance. The model can accept both text and images and generate text, with a context window of up to one million tokens. However, the actual context window may vary depending on factors such as runtime state, serving framework, other applications, and chosen quantization.\n\nAlthough the model is open-weight, released under the Apache 2.0 license, privacy ultimately depends on the entire system setup. A coding agent with web search, remote APIs, telemetry, cloud storage, or unrestricted shell access could still expose data outside the machine. Independent evaluations, such as the Intelligence Index score of 20 reported by Artificial Analysis, suggest that Inkling 2 Mini performs at a comparable level to other AI models like Qwen3.8 27B and Nanbeige4.3-33B. However, these comparisons should be taken cautiously, as they do not directly measure the energy use or latency of a user's specific laptop.\n\nInkling 2 Mini's performance varies across different benchmarks, as seen in its mixed results on Terminal Bench v4.0 and Terminal Bench v2.1. The model card highlights that looping does not simply add factual knowledge but rather treats plausible explanations and generated code as hypotheses that need to be checked, especially when dealing with facts beyond the scope of the prompt. When using Inkling 2 Mini, it is essential to evaluate the model's performance on specific tasks using a small set of tasks you are familiar with, such as explaining a function, writing tests for an existing bug, and proposing a patch in a disposable checkout. Be sure to record the runtime, memory consumption, and test results, and inspect the differences yourself. Additionally, if connecting an agent, limit filesystem and network permissions and avoid providing secrets by default.",
  "summary": "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…",
  "key_points": [
    "Inkling 2 Mini is a locally runnable multimodal model.",
    "Loop design increases computation without increasing weights.",
    "Model performs at level comparable to Qwen3.8 27B and Nanbeige4.3-33B."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}