← Back to BLACKWIRE VOLT BUREAU AI PERFORMANCE Screenshot of Inkling 2 Mini building a flight simulator on a MacBook Pro.

Inkling 2 Mini generated a playable flight simulator in minutes on a consumer laptop during a flight test.

THINKING MACHINES LAUNCHES INKLING 2 MINI, TOP OPEN-WEIGHTS MODEL <100B

*Inkling 2 Mini arrives on Nov 20, 2026 as a looped transformer small enough for laptops. It outperforms every open‑weights model under 100 B parameters on key coding benchmarks while keeping a 20 GB memory footprint.*

By VOLT Bureau - BLACKWIRE  |  November 21, 2026, 03:00 CET  |  Inkling 2 Mini, looped transformer, open‑weights AI, laptop AI models

Inkling 2 Mini debuted on Nov 20, 2026 as a looped transformer from Thinking Machines Lab. The architecture runs the middle layers three times per token, delivering richer reasoning while keeping active parameters to 3.4 B out of a 26.4 B total. Its ~20 GB memory footprint fits on consumer laptops, making high‑end AI accessible outside data‑center clusters.

In head‑to‑head tests Inkling 2 Mini outstrips every open‑weights model under 100 B parameters on DeepSWE v1.1 (48.2 %) and Vals Index v2.2 (46.5 %). It also runs faster than similarly scoring models, and its open‑weight release on Tinker invites immediate experimentation by researchers and developers.

What makes Inkling 2 Mini’s looping architecture different?

Inkling 2 Mini is a Mixture‑of‑Experts model that loops its middle transformer layers three times per token. The design activates only 3.4 B of its 26.4 B total parameters, but requires 5.4 B weight activations per token to support the loops. Looping lets the model extend its reasoning horizon without writing intermediate thoughts, improving chain‑of‑thought (CoT) monitorability. A variable‑loop variant showed the no‑CoT time horizon grow from 16 s at one pass to 37 s at three passes, and performance on a serial‑reasoning benchmark kept rising through ten loops. The CoT remained legible even as loops increased, and sabotage scores did not rise.

Benchmark dominance under 100B parameters

Inkling 2 Mini leads every open‑weights model with fewer than 100 B parameters on DeepSWE v1.1 (48.2 %) and Vals Index v2.2 (46.5 %). It also tops the field on agentic coding benchmarks such as SWE‑bench Pro v2, falling behind only Qwen3.8‑27B on a few other tests. On the kernel‑optimization benchmark PMPP‑Hard, its score approaches that of the much larger Inkling 2 Small. While looping boosts reasoning, factuality measured by AA‑Omniscience matches peers of the same size, confirming that knowledge remains tied to total parameter count.

We’re excited to offer some of the first models with looped architecture available to researchers.

Laptop‑ready performance and real‑world demo

The model’s memory footprint sits at roughly 20 GB, enabling inference on consumer laptops and desktop AI rigs. Researchers ran Inkling 2 Mini offline on a MacBook Pro (M5 Pro, 48 GB) during a flight, building a playable flight simulator in minutes. Decoding speed hits a theoretical ceiling of 614 GB/s with BF16 weights and batch size one, even with repeated reads across loops. Most weights are deployed natively in NVFP4, with the final four layers in MXFP8, ensuring efficient hardware utilization. The open‑weights release on Tinker allows immediate customization and testing.

Limitations: knowledge vs reasoning

Looping improves long‑context reasoning but does not expand the model’s factual store. Benchmarks that stress pure knowledge—AA‑Omniscience, SimpleQA Verified, SimpleVQA, WorldVQA—show no gains from additional loops. Conversely, long‑context tasks like GraphWalks BFS and MRCR v2 improve as passes increase. Inkling 2 Mini compensates for its knowledge gap by leveraging search and tool use, leading open‑weights peers on BrowseComp, DeepSearchQA, and MCP Atlas. The pattern underscores that reasoning depth and factual breadth remain distinct engineering challenges.

Frequently asked questions

How does Inkling 2 Mini’s looping affect its reasoning ability?

Looping runs the middle transformer layers three times per token, extending the model’s reasoning horizon without increasing parameter count, which improves performance on long‑context benchmarks and chain‑of‑thought monitorability.

Can Inkling 2 Mini run on a typical laptop?

Yes. Its memory footprint is about 20 GB, and researchers ran it offline on a MacBook Pro (M5 Pro, 48 GB) to build a flight simulator in minutes.

What benchmarks does Inkling 2 Mini lead among open‑weights models under 100B?

It leads on DeepSWE v1.1 (48.2 %) and Vals Index v2.2 (46.5 %), and outperforms peers on agentic coding tests such as SWE‑bench Pro v2, while trailing only Qwen3.8‑27B on a few other tasks.

Inkling 2 Mini proves that looped transformers can deliver top‑tier reasoning on modest hardware, reshaping the economics of AI deployment. Yet its factual limits remind the community that scaling knowledge still demands larger parameter counts or smarter retrieval. As developers integrate the model into tools and workflows, the industry will watch whether looping becomes a standard design for compact, high‑performance AI.

Sources: Hacker News – https://thinkingmachines.ai/news/inkling-2-mini/