Skip to content
← Back to Price per IQ

A 20 GB open model for laptops outscores the same lab's 975B model from July

Thinking Machines' Inkling 2 Mini matches a 27B rival on Artificial Analysis while using about 2.8× less compute per task.

20Artificial Analysis score, against 12.5 for the 975B Inkling (July)

On 20 November 2026 Thinking Machines released Inkling 2 Mini with open weights (anyone can download and run the model) under an Apache 2.0 licence. It has 26.4 billion parameters in total, of which 3.4 billion are used for each token, and the checkpoint is about 20 GB. The lab’s demo ran it on a MacBook Pro with an M5 Pro chip and 48 GB of memory, and the Hugging Face card recommends 32 GB or more.

Artificial Analysis, an independent testing group, scored it 20 on its Intelligence Index, level with Qwen3.8 27B and Nanbeige4.3-33B. It reports that Inkling 2 Mini needs about 2.8× less compute than Qwen3.8 27B, but writes about 152,000 output tokens per task against about 110,000. OpenRouter’s model list carries the same index for the lab’s earlier models: Inkling, its 975 billion parameter model from July 2026, shows 12.5, and Inkling 2 Small, released on 7 November, shows 30.6.

Inkling 2 Mini is a looped model: its middle 12 layers run three times with shared weights, which the model card says brings the work per token to about 5.4 billion parameters. OpenRouter prices it at $0.10 per million input tokens and $0.40 per million output tokens.

What it means

A level of ability that took a 975 billion parameter model in July now fits in a download a laptop can hold. This is the pattern of the cost-of-intelligence charts: a fixed level gets cheaper and smaller, while the best models stay large.

Caveats

  • The laptop claim needs a laptop with 32 GB of memory or more, which is a high-end machine. Thinking Machines does not say how fast it runs there.
  • Thinking Machines reports the headline scores of 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2 itself, and notes that “Evaluation setups vary by model”. Only the Artificial Analysis score is independent.
  • The scores for the older Inkling models come from OpenRouter’s listing, read on one day, which does not name the index version. The site’s Artificial Analysis chart uses version 5.0.1, so these scores are not placed on it.
  • Looping and extra output tokens mean the 3.4 billion active parameters understate the work done per answer.
  • Thinking Machines says it leads open models under 100 billion parameters on some tests, but its own table has Qwen3.8 27B ahead on Terminal Bench v2.1 (73.0% against 71.6%).
Suggest a correction