Kivorent News

Sunday, November 22, 2026

Tech and AINews

Thinking Machines' Inkling 2 Mini is an open-weight model that fits on a 32 GB laptop

Thinking Machines Lab released Inkling 2 Mini on November 20, a 26.4-billion-parameter open-weight model under Apache 2.0 that runs in about 20 GB. Artificial Analysis scores it level with Qwen3.8 27B at a sixth of the cost per task.

Cortex, covering AI models, labs and agents
, 3 min read

~20 GBInkling 2 Mini's native checkpoint, vs 394 GB for Small

What to know

  • Thinking Machines Lab released Inkling 2 Mini's weights on November 20 under an Apache 2.0 license with an acceptable use policy.
  • The model has 26.4 billion parameters and runs its middle 12 layers three times per token, a design called a looped transformer.
  • Artificial Analysis gives it 20 on its Intelligence Index, matching Qwen3.8 27B at $0.59 per task against $3.53.
Cost to run Artificial Analysis's index tasks, per task
Inkling 2 Mini$0.59
Qwen3.8 27B$3.53

Source: Artificial Analysis, November 20, 2026. Both models score 20 on its Intelligence Index.

Eight days ago, running Thinking Machines' best open model meant renting a multi-GPU server: Inkling 2 Small's weights came to 394 GB. The lab's new Inkling 2 Mini comes in at about 20 GB, roughly a twentieth of that, and VentureBeat reports it runs offline on a 32 GB MacBook Pro or a single desktop GPU.

Mira Murati's lab posted the weights on November 20 under Apache 2.0, the same permissive license as Small, with an acceptable use policy attached, according to Artificial Analysis and LM Studio. The model takes text and images, writes text, and has a 1-million-token context window, though LM Studio warns that the smallest memory setups can't hold anything close to the full context.

Three passes through the middle

The interesting part is the architecture. Inkling 2 Mini is a mixture-of-experts model with 26.4 billion total parameters across 24 stored layers. The first six and last six layers run once; the middle 12 run three times, reusing the same weights, so each token goes through 48 layer applications, per VentureBeat. That's the "looped" in looped transformer: it buys extra computation per token without storing more parameters.

It also muddles the usual size numbers. VentureBeat, TestingCatalog and LM Studio give 3.4 billion active parameters per pass and 5.4 billion counting the loops. Artificial Analysis also uses 5.4 billion but says only 3.1 billion unique weights touch each text token. Nobody has reconciled the 3.4 and 3.1, and the lab's own announcement and model card were not readable for this story.

What the independent numbers say

Thinking Machines' model card claims 48.2% on the DeepSWE v1.1 coding benchmark and 79.2% on SWE-bench Verified, ahead of Alibaba's dense Qwen3.8 27B on both, according to LM Studio's copy of the card. Those are the lab's runs; see what a benchmark score does and does not tell you.

Artificial Analysis, which ran its own tests, is less flattering and more useful. It scores Inkling 2 Mini 20 on its Intelligence Index, tied with Qwen3.8 27B, against 31 for Inkling 2 Small. The tie hides a real difference in cost. Inkling 2 Mini spent about 38% more output tokens to get there, around 152,000 per task, yet finished each task in 9 minutes against 38.2 and cost $0.59 against $3.53, about six times cheaper. The lab's own API charges $0.15 per million input tokens and $0.50 per million output, exactly half of Small's $0.30 and $1.00.

The weak spots are listed in plain view. On the AA-Omniscience factual-recall test it scores -8.6, worse than Qwen's -6.4, and it trails Qwen on multi-turn chat and on instructions with many constraints. VentureBeat and LM Studio both mention repetitive reasoning while debugging.

The monitorability claim

Thinking Machines is selling the loops partly as a safety result. In preliminary tests reported by VentureBeat and TestingCatalog, looping more than doubled the model's reasoning horizon without a written chain of thought, from 16 seconds at one pass to 37 at three, and cut chain-of-thought length by more than half. The lab says the traces that remain are still legible and found no rise in covert sabotage on the LinuxArena benchmark. Murati told TestingCatalog that the model's thinking "remains monitorable; this may not stay the case."

That sentence carries the story. A model that does more of its thinking where nobody can read it is cheaper to run and harder to check, and the lab is the one saying the trade-off may get worse. Once weights are out, as we noted with Small, nobody can pull them back. For the wider trade-offs, see open-weight versus closed AI models. The release lands as the lab is reported to be raising $5 billion at a $40 billion valuation.

Sources

AI modelsAI agents

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.