EXO Labs says it has made Thinking Machines’ Inkling 2 Mini run at 210 tokens a second on a 96GB M5 Ultra Mac Studio, in a post on X on Saturday. We covered the model’s launch in our Inkling 2 Mini piece.
On a 48GB M5 Max MacBook Pro, EXO reports 125 tokens a second. A token is a word fragment, so this is roughly how fast the model writes an answer on a desk or a laptop, with no cloud server.
Business Pill · Free Cash Flow
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this model.
The key insight: EXO Labs reports Inkling 2 Mini at 210 tokens a second on a 96GB Mac Studio and 125 on a 48GB MacBook Pro, about 4.8 and 8.0 seconds for a 1,000-token answer (our arithmetic).
Two Speeds: Reading and Writing
EXO gives two numbers per machine. Prefill is how fast the model reads the prompt before it answers. Decode is how fast it writes new tokens. Every score and price is tracked on our AI model benchmarks page.
On the Mac Studio, EXO reports 3,620 tokens a second for prefill and 210 for decode. On the MacBook Pro, it reports 1,750 and 125.
The Mac Studio decodes 1.68 times as fast and prefills 2.07 times as fast as the laptop, with twice the memory (our arithmetic).
At 210 tokens a second, a 1,000-token answer takes about 4.8 seconds; at 125, about 8.0 seconds (our arithmetic).

The Claude Comparison
EXO says the model “built a native weather app 1.8x faster than Claude Sonnet 5.5 on an M5 Ultra Mac Studio and 11x cheaper (API pricing).”
That is EXO’s own test. Julien Chaumond quoted the line on X and added: “And this, my friends, is how local and open source AI is going to win”, in his post.
Why Looping Suits a Laptop
Inkling 2 Mini runs its middle 12 of 24 layers three times per token with shared weights, per its model card. EXO says “looping layers of the model increases inference-time compute (and intelligence) without increasing the size of the model.”
EXO calls the design “ideal for local devices like MacBooks, which are memory constrained.” The checkpoint is about 20GB, per the model card.
The Structural Read
As we read it, the claim that matters is that a 20GB open model now writes fast enough on one Mac to be usable for coding.
As we read it, the Claude comparison is EXO’s own test until it publishes the method.
EXO Labs, 21 Nov 2026
“Looping layers of the model increases inference-time compute (and intelligence) without increasing the size of the model.”
Three Implications
FOR DEVELOPERS EXO reports the model at 125 tokens a second on a 48GB MacBook Pro.
FOR CLOUD PROVIDERS EXO says its test was 11x cheaper than Claude Sonnet 5.5 at API pricing; the method is unpublished.
FOR BENCHMARKS the speeds are EXO-reported; independent tests are not out yet.
What Is Not Established
EXO has not published how it ran the weather-app test: the task, the number of runs, how it timed “faster”, or which API prices it used. We have not seen independent tests of these speeds.
The Bottom Line
EXO Labs reports 210 tokens a second for Inkling 2 Mini on a Mac Studio and 125 on a MacBook Pro, and says its weather-app test beat Claude Sonnet 5.5 on speed and cost.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The speed figures and the Claude Sonnet 5.5 comparison are EXO Labs’ own, from its post on X; EXO has not published the test method. Model details are from Thinking Machines’ model card via our launch piece. The ratios and answer times are our arithmetic.
Sources: EXO Labs post on X, 21 Nov 2026 01:37 UTC · Julien Chaumond post on X, 21 Nov 2026 11:17 UTC · FourWeekMBA, 20 Nov 2026: Inkling 2 Mini launch and model card









