EXO Labs Says It Ran Thinking Machines Lab’s Inkling 2 Mini at 210 Tokens a Second on a Mac Studio; the Figures Are EXO’s Own and Its Engine Is Not Yet Released

EXO Labs says it worked with Thinking Machines Lab to make the open-weights model Inkling 2 Mini run fast on Macs: 210 tokens a second of decode and 3,620 of prefill on a 96GB M5 Ultra Mac Studio, and 125 and 1,750 on a 48GB M5 Max MacBook Pro, according to EXO’s post, which Lens received as text copied from LinkedIn and whose opening lines VentureBeat quotes from a version on X. The figures are EXO’s own. The inference engine behind them is described as upcoming and unreleased, and EXO has not stated the quantization, context length or batch size. One independent test on a 96GB M5 Ultra, using different software, measured 83.4 tokens a second at 4,000 tokens of context. Thinking Machines gives no tokens-per-second figure for Macs. Lens covered the company’s funding in an earlier post.
EXO’s video
EXO Labs’ post on X, shown below with its video, is the source of the opening lines quoted in this story: “We worked with @thinkymachines to make Inkling 2 Mini run fast locally,” followed by the speeds above, as VentureBeat reproduces it. The video belongs to EXO Labs and is shown through X’s own embed; Lens has not been able to watch it and does not describe what it shows or check its claims against the footage. Watch it on X: x.com/exolabs/status/2123953913342935486/video/1.
Video: EXO Labs (@exolabs) on X, via X’s embed.
What EXO says
- Speeds. 210 tokens a second decode and 3,620 prefill on a 96GB M5 Ultra Mac Studio; 125 and 1,750 on a 48GB M5 Max MacBook Pro.
- A weather app. 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).” It does not say how the speed or cost was measured. Claude Sonnet 5.5 is a model from Anthropic, the maker of Claude.
- The timeline. EXO says it got access on Wednesday, had the model running at 30 tokens a second in three hours and at 210 a day later, “7x faster” (Lens arithmetic: 210 divided by 30 is 7). It adds: “The Thinking Machines team couldn’t believe it.”
- How. EXO calls this “one of the first real tests of our upcoming Apple Silicon inference engine and agent pipeline that continuously optimizes inference across our fleet of Macs.” It says it gave the system no guidance about the looping architecture, and that the system “built a working implementation and found architecture-specific optimizations on its own,” improving kernels and fusions and making better use of DSpark with adaptive drafting.
- Why looping suits laptops. EXO says looping layers increases inference-time compute “without increasing the size of the model,” which suits memory-constrained devices.
What Thinking Machines released
- The model. Inkling 2 Mini was released on 20 November 2026 with open weights on Hugging Face under Apache 2.0, according to Thinking Machines’ model card; VentureBeat adds an acceptable use policy. It has 26.4 billion parameters in total, 3.4 billion active per token and 5.4 billion counting the repeated passes, in 24 layers whose middle 12 run three times with shared weights, giving 48 layer applications per token. Its checkpoint is about 20GB, and the context window goes up to 1 million tokens.
- DSpark. Hugging Face’s launch post says “a DSpark drafter is also included for additional speed-up at a small memory cost.” That is the only page Lens read that explains the term.
- On Macs. The model card lists Apple silicon devices with 32GB or more of memory, running through Ollama and EXO Labs, and says “speed varies by runtime and hardware.” Thinking Machines’ announcement shows the model building a flight simulator offline on a MacBook Pro with an M5 Pro and 48GB, but gives no tokens-per-second figure; its decoding chart is a theoretical ceiling at 614 GB/s of memory bandwidth.
- Benchmarks. Thinking Machines reports 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2 and says the model leads open-weights models under 100 billion parameters on benchmarks such as DeepSWE, though its own table compares four models. These are the company’s results.
- Limits it states. The model card says the model sometimes becomes stuck in long reasoning loops, particularly when debugging code, and shows clipped phrasing in Chinese.
- Price and hosted speed. VentureBeat lists $0.15 per million input tokens and $0.50 per million output tokens at Thinking Machines, and cites Artificial Analysis for 280 output tokens a second on the hosted service; Thinking Machines’ own pages give no prices.
Other speed measurements
These use different hardware, software and settings, so they are not like-for-like with EXO’s claim.
- RigBase (DEV Community, 21 November) lists, as independent, a 96GB M5 Ultra running oMLX with a multi-token-prediction setting at 83.4 tokens a second of decode and 3,225 of prefill at 4,000 tokens of context, and 56.2 and 2,968 at 32,000; a 512GB M3 Ultra at 40.3 and 1,169 at 1,000 tokens; and a 64GB M5 Pro Mac mini on llama.cpp at 29.95 and 490.8 for a 512-token prompt. EXO’s figure is about 2.5 times the first of these (Lens arithmetic). RigBase says it did not re-run EXO’s numbers, calls them vendor-reported, and says EXO’s blog and benchmark site had no Inkling write-up on 21 November.
- Kingy measured a median of 11.6 tokens a second on a 24GB M4 Pro MacBook Pro using a 2-bit quantization with some experts on the CPU.
- A developer, @KyleHessling1, reported 104 tokens a second on an Nvidia RTX 5090, according to VentureBeat. It is not a Mac.
Where reports differ
- Memory. The model card says Apple silicon with 32GB or more; VentureBeat’s table also lists 16GB or more through GGUF files and quotes Unsloth saying the model runs on 16GB of RAM.
- Parameter counts. Hugging Face rounds to about 26 billion in total and 3 billion active; Thinking Machines gives 26.4 and 3.4 billion.
- The weather app. The sources Lens read give no basis for the 1.8 times and 11 times figures, and VentureBeat’s price table does not include Sonnet 5.5, so Lens could not reproduce the cost comparison.
Sources and notes
Lens read EXO’s post as text copied from LinkedIn and pasted to it, Thinking Machines’ announcement and model card, Hugging Face’s launch post, VentureBeat, 21 November (by Carl Franzen, which quotes EXO’s post on X), RigBase, Kingy and MLQ, through automated page summaries. MarkTechPost and FourWeekMBA returned 403 errors. X would not open EXO’s post for Lens, so Lens has not seen the video and relies on the link supplied to it (the post’s /video/1 address) for the fact that the post carries one; the longer text, including the timeline, comes from the pasted copy. Lens did not run any test, and found no Thinking Machines response to EXO’s figures.
Image: a Mac Studio with a Studio Display, keyboard and trackpad in an Apple Store, a file photo used as a stand-in. It does not show the machine EXO tested. Photo by AzureSaturn, CC0, via Wikimedia Commons.

