Thinking Machines' Open-Source Inkling 2 Mini Matches Frontier Models at a Fraction of Cost

Thinking Machines Lab (TML) released the open-source small model Inkling 2 Mini in November 2026, with 26.4B total parameters and 3.4B activated (effectively around 5.4B). It runs locally on machines with roughly 16–20GB of VRAM/memory or 32GB of RAM. Results from multiple third-party evaluation organizations show that, with a tiny parameter count and cost, the model matches or even outperforms flagship models with an order of magnitude more parameters on several benchmarks, making it the center of this round of discussion.

Confirmed

Unconfirmed

Why it matters

Inkling 2 Mini demonstrates a route opposite to "stacking parameters": using architectural efficiency (such as recurrent layers) to approach flagship-level intelligence on consumer hardware, while cutting per-task costs to a fraction of the original and minimizing energy consumption. If the data holds up under further independent verification, it could reshape developers' expectations of small models' cost-effectiveness in replacing frontier models for coding, cybersecurity, and automation pipelines—and push "intelligence per joule" as a new model evaluation metric.

2026-11-20 ~ 2026-11-21 · 20 related posts

Full story(16 episodes)→

Primary sources

1 near-duplicate retellings: Aubrey_D_Graham