Inkling 2 Mini Brings Looped AI Model to Local Development
Thinking Machines Lab releases Inkling 2 Mini, a locally runnable multimodal AI model with a unique looped architecture for coding and tool use.

CANADA — Thinking Machines Lab has launched Inkling 2 Mini, an open-weights multimodal AI model designed to run locally on developer machines. The model, released on November 20, features a unique looped architecture that reuses internal layers three times per token, allowing for more computational power without significantly increasing stored weights. This design makes it particularly suited for tasks like coding and tool use, though developers should note that 'runs on a laptop' doesn't guarantee optimal performance on all hardware.
With 26.4 billion total parameters, Inkling 2 Mini operates as a mixture-of-experts model, routing each token to a subset of experts rather than using all for every token. The model card details 3.4 billion active parameters and approximately 5.4 billion effective active parameters per token after accounting for the loop passes. The architecture includes 24 physical decoder layers, with the middle 12 layers running three times, effectively creating 48 layer applications using only 24 layers of weights.
While the model's looped design saves storage, it doesn't eliminate computational costs. The official checkpoint is about 20 GB in a mixed NVFP4 and MXFP8 format, requiring substantial hardware. Thinking Machines recommends Apple silicon computers with at least 32 GB of unified memory or NVIDIA DGX Spark for local operation. The model accepts both text and images, producing text outputs with a context window of up to one million tokens, though real-world performance will vary based on hardware and workload.
Released under the Apache 2.0 license, Inkling 2 Mini offers developers the option for local inference, but privacy depends on the entire setup. The model has shown promising results in benchmarks, scoring 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified in specific configurations. However, developers should treat its outputs as hypotheses to verify, especially for fact-dependent tasks. The model is available via Hugging Face and supports various runtimes, including Ollama for Apple silicon and llama.cpp for Linux, though developers should check documentation for specific hardware compatibility.
NewsCritical Check Point Vulnerability CVE-2026-123039 Exploited in PHP2Shell Campaign
Security researchers detail post-exploitation evidence and recovery steps for the exploited Check Point CVE-2026-123039 vulnerability with CVSS 7.8 score.
NewsDeveloper Reveals Method for Scraping App Store Reviews Without Proxies
A Node.js developer shares a technique for extracting international App Store review data in JSON format, enabling sentiment analysis without complex scraping infrastructure.
NewsDatabricks SQL AI Forecast V2 Adds Holiday Support and Covariates
Databricks SQL's ai_forecast function now supports holiday effects and covariates, streamlining time-series forecasting directly within SQL queries.
NewsHow a WordPress LMS Front End Was Rebuilt with React and Cloudflare Pages
A team modernized a WordPress LMS front end using React and Cloudflare Pages while maintaining core functionality like logins and payments.