AI Trends

Thinking Machines Releases Inkling 2 Mini, an Open-Weight Recurrent Transformer Model

Full Article from Artificial AnalysisUpdated Nov 21 04:53

AI summary

Thinking Machines has released Inkling 2 Mini, which scored 20 on the Artificial Analysis Intelligence Index and reached the Pareto frontier for intelligence versus parameter count. The model uses a MoE architecture with 26B total parameters and 5.4B active parameters, supports text and image inputs and a 1M-token context window, and is available as open weights under the Apache 2.0 license. It matches the score of Qwen3.8 27B (xhigh), while being about 6x cheaper and 4x faster per task, but producing approximately 38% more output tokens.

Why read this

SelectedAI score 73

Comparative data on parameters, compute, cost, and speed illustrates the efficiency trade-offs of the recurrent-layer design in small models.

Topics

Tags

Body

Artificial Analysis

All articles

November 20, 2026Thinking Machines has released Inkling 2 Mini, an open weights looped transformerSee model page

Thinking Machines has released Inkling 2 Mini at 20 on the Artificial Analysis Intelligence Index, an open weights looped transformer on the Pareto frontier for Intelligence vs. Active Parameters

Inkling 2 Mini is a 26B total, 5.4B active parameter mixture-of-experts model. The 5.4B active parameter count includes three passes through its looped layers; it uses 3.1B unique weights per text token, and it is small enough to run on a MacBook or DGX Spark.

It is released under an Apache 2.0 license, with text and image input and a 1M token context window.

Inkling 2 Mini matches Qwen3.8 27B (xhigh) and Nanbeige4.3-33B on the Intelligence Index with ~5x and ~6x fewer active parameters, respectively.

It is on the Pareto frontier for Intelligence vs. size on both total and active parameters.

Inkling 2 Mini uses ~38% more output tokens than Qwen3.8 27B for the same Intelligence Index score, but ~2.8x less compute.

Twelve of its 24 layers run three times per token, adding compute without adding output tokens. It averages ~152K output tokens per Intelligence Index task vs. ~110K for Qwen3.8 27B and ~158K for Nanbeige4.3-33B. On our compute proxy measure (active parameters × tokens), it sits on the Pareto frontier, using ~2.8x less compute than Qwen3.8 27B and ~5x less than Nanbeige4.3-33B.

Inkling 2 Mini is ~6x cheaper and ~4x faster per task than Qwen3.8 27B.

It costs $0.59 and takes 9.0 minutes per Intelligence Index task vs. $3.53 and 38.2 minutes for Qwen3.8 27B. GPT-6 Luna (max) scores 28 for $0.15 per task. Cost uses Thinking Machines' first-party pricing ($0.15/$0.50/$0.015 per 1M input/output/cache hit tokens).

Full Intelligence Index evaluations breakdown:

Google continues to accelerate its release cadence with Gemini 4 Neon

Google continues to accelerate its release cadence with Gemini 4 Neon, delivering a small Intelligence gain over Gemini 4 Carbon at a ~17% lower Cost per Task

November 18, 2026

OpenAI has the #2 model in the Intelligence Index with the release of GPT-6.2 Astra and GPT-6.2 Sol

OpenAI’s release of GPT-6.2 Astra and GPT-6.2 Sol delivers the #2 model in the Artificial Analysis Intelligence Index and most of the Cost per Task Pareto frontier

November 17, 2026

Multimodal Knowledge Agent (MKA) benchmark: testing knowledge work capabilities with complex visual files

We’ve partnered with Google DeepMind to launch their Multimodal Knowledge Agent (MKA) evaluation, a new benchmark for agentic knowledge work over complex visual files

November 10, 2026

Source

Full Article from Artificial Analysis