AI News
All stories
Reported

Thinking Machines releases open-weights Inkling 2 Mini, a small looped transformer

Thinking Machines has released Inkling 2 Mini under Apache 2.0, and Artificial Analysis gives it a score of 20 on its Intelligence Index with 5.4B active parameters.

Models · · 2 min read

Thinking Machines has released Inkling 2 Mini, an open-weights model that Artificial Analysis says scores 20 on its Intelligence Index. The model activates 5.4B of its 26B total parameters per token. It uses a mixture-of-experts design, is released under an Apache 2.0 license, accepts text and image input, and offers a 1M token context window. Artificial Analysis says it sits on the Pareto frontier for intelligence against both total and active parameter counts. It matches Qwen3.8 27B (xhigh) and Nanbeige4.3-33B on the index with roughly 5x and 6x fewer active parameters, respectively.

Twelve of its 24 layers run three times per token, so the 5.4B active parameter figure includes three passes through the looped layers. Per text token, the model uses 3.1B unique weights, which Artificial Analysis says is small enough to run on a MacBook or DGX Spark. The looping adds computation without adding output tokens.

Verbosity is the other side of those numbers, and Artificial Analysis's figures show it. Inkling 2 Mini averages about 152K output tokens per Intelligence Index task. Qwen3.8 27B averages about 110K, and Nanbeige4.3-33B about 158K. Artificial Analysis says this is about 38% more output than Qwen3.8 27B for the same score, but about 2.8x less compute, using its own compute proxy of active parameters multiplied by tokens. Its evaluation generated 161M output tokens in total, against a median of 85.7M for comparable open-weights models. The site describes that as very verbose.

On cost and speed, the firm's figures favor the Thinking Machines model over Qwen3.8 27B. Inkling 2 Mini costs $0.59 and takes 9.0 minutes per Intelligence Index task. Qwen3.8 27B costs $3.53 and takes 38.2 minutes. Artificial Analysis describes the model as about 6x cheaper and about 4x faster per task. Those cost figures use Thinking Machines' first-party pricing: $0.15 per 1M input tokens, $0.50 per 1M output tokens, and $0.015 per 1M cache-hit tokens. The comparison has caveats. Artificial Analysis notes that GPT-6 Luna (max) scores 28 for $0.15 per task. The listed price is also only one provider's rate, since the site counts the model as available through one API provider.

What changed

Thinking Machines released Inkling 2 Mini as an open-weights model under Apache 2.0, with 26B total and 5.4B active parameters per token and a 1M token context window.

Type: New capability · Applies to: Inkling 2 Mini

Why it matters

Teams that want to run a capable model on a laptop or small server get an openly licensed option. Artificial Analysis rates it as cheaper and faster per task than Qwen3.8 27B at a similar score. The figures are Artificial Analysis's own benchmark results.

What you can do

Test Inkling 2 Mini on your own workloads against your current small open-weight models before changing anything. The comparison rests on one benchmark index.

Original source

We have no link to the original source for this story yet.

Further reporting

Where we saw it

1 source

Artificial Analysis

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works