Inkling 2 Mini, a local model betting on loops
Thinking Machines Lab has introduced Inkling 2 Mini, an open weights model designed for local use. Its uniqueness lies in a looped architecture that increases computing capacity without significantly increasing the number of stored weights.
Thinking Machines Lab unveiled Inkling 2 Mini on November 20, 2026, a multimodal open weights language model designed to run locally on consumer machines or workstations. Its uniqueness lies in an architecture with repeated passes in the central layers, intended to improve reasoning without drastically increasing the model size. This topic matters for off-cloud uses, but also for how model quality is now measured: not just by size, but by the way it “thinks.”
Une architecture qui boucle sur elle-même, sans tourner en rond
Inkling 2 Mini is a Mixture-of-Experts model, meaning only certain internal experts are activated for each token processed. Here, the 24 layers are organized into four blocks of six: the first and last pass once, while the two middle blocks are replayed three times with the same weights. Result: 48 layer visits per token, for only 24 physical layers.
The technical promise is simple to state, harder to train. Successive passes exchange messages via control mechanisms that regulate what the layer should retain or forget. Thinking Machines Lab explains that this setup allows adding “logical” depth without storing more parameters. In other words, the model works harder but hasn’t grown so much as to require a club chair.
Des chiffres pensés pour le local, pas pour les serveurs géants
The model has about 26.4 billion total parameters, with 3 billion active parameters per token according to the launch summary, and a memory footprint of about 20 GB. Deployment variants include a mixed format with expert weights in NVFP4 and MXFP8, as well as support for common local stacks: transformers, llama.cpp, MLX, and SGLang. Thinking Machines Lab also provides a research variant with variable pass count, TML-d24-varloop-research [to be verified for exact production use].
Multimodal compatibility is also included: text and images can be injected in the same request. This is practical for desktop, support, or prototyping uses, where one might want to ask a model to describe a screenshot, comment on a diagram, or read an interface without going through a separate pipeline.
Pourquoi cette approche intéresse autant les chercheurs que les intégrateurs
Beyond the technical sheet, Inkling 2 Mini touches on a fundamental debate: can reasoning be improved without multiplying parameters and inference costs? Provided sources indicate the model achieves good results on several benchmarks, notably in code and tool-assisted search, outperforming several competitors in many benchmark categories. Another source cites 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 Hard, versus 42.2% and 73.8% for Qwen3.8-27B [to be verified according to exact test protocol].
Thinking Machines Lab also advances a more strategic argument: looping would improve reasoning without allowing weights to “memorize” more facts. In short, the architecture mainly helps better organize computation, not artificially inflate the knowledge base. For integrators, the conclusion is immediate: if performance holds, a model of this format could become a credible option for local deployments, including on machines with 32 GB RAM or modest servers.
Les démonstrations veulent convaincre sur un point: ça tourne sans le cloud
The launch relies on several demonstrations. Thinking Machines Lab shows Inkling 2 Mini on a MacBook Pro in airplane mode, where the model generates a flight simulator in minutes, then a “pixel town” where twenty persistent inhabitants are managed locally. The message is clear: the machine can stay alone in the cabin. And it doesn’t even need Wi-Fi to pretend to understand everything.
The company also publishes examples of WebGL rendering, reconstruction of a Van Gogh painting, or expense management application, illustrating the model’s versatility on coding and multimodal tasks. These demonstrations don’t replace independent evaluations but give an idea of the targeted field: local AI capable of chaining vision, code generation, and tool-assisted reasoning without relying on a remote service.
Ce que ce lancement dit du marché des modèles ouverts
Inkling 2 Mini arrives at a time when open models aim less to break size records than to gain efficiency, portability, and usage cost. The positioning is clear: a model compact enough to run locally, yet sophisticated enough to serve as a base for agents, code assistants, or business tools. For publishers and technical teams, it’s a concrete path; for researchers, it’s an observation ground on the relationship between computational depth, monitorability, and reasoning quality.
One important limitation remains: the monitorability and reasoning results provided by Thinking Machines Lab are still preliminary, and some comparisons strongly depend on the chosen protocol. But the launch still marks a clear milestone: innovation no longer passes only through bigger models, but also through architectures that do more with fewer stored parameters.