
Share
A 26.4B-parameter looped transformer from Mira Murati's startup matches a 27B dense rival on intelligence benchmarks while running four times faster and six times cheaper, and it fits entirely on consumer hardware.
Thinking Machines Lab just released a model that's small enough to run offline on a MacBook Pro but scores like something far bigger. Inkling 2 Mini, dropped yesterday under an Apache 2.0 license, is a 26.4-billion-parameter multimodal model with a roughly 20GB memory footprint. It runs on a 32GB MacBook Pro or a single desktop GPU, no cloud required. And on software engineering and tool-use benchmarks, it holds its own against models nearly its size and some considerably larger.
The interesting part isn't the parameter count. It's how Thinking Machines got reasoning performance out of a model this small without just piling on more weights.
Standard transformers process tokens in one linear pass, first layer to last. When a problem needs more sequential reasoning than the model's depth can provide, it has to externalize that thinking as long chains of thought in text. That's expensive: more output tokens, more latency, higher inference cost. Inkling 2 Mini sidesteps this with what Thinking Machines calls a "looped transformer" architecture.
Per the model card, it's a 24-layer, decoder-only sparse Mixture-of-Experts model routing each token to 8 of 128 experts, activating 3.4 billion parameters per standard pass. Here's the twist: the first six layers and the last six run once, but the middle 12 layers run three times using shared weights. Picture an engineer triple-checking a tricky equation mentally before committing to the next line of code, rather than filling three pages with scratch work.
Those three passes produce 48 effective layer applications per token and 5.4 billion effective active parameters, with no increase in storage size. Thinking Machines uses attention residuals (a technique letting each pass read a gated "message" from the prior pass's middle layers) to keep information coherent across iterations. In internal scaling tests, a non-looping baseline needed 140% more inference compute to match the looped model's loss on 512K-token agentic traces. That's a meaningful compute savings for anyone running this at scale.
Getting a capable MoE model down to laptop-friendly size took some surgery. Thinking Machines pruned the expert count from 384 down to 128 mid-training, then "healed" the resulting network by distilling 1 trillion tokens from a larger internal Inkling model, followed by supervised fine-tuning and 2.2 trillion tokens of reinforcement learning.
The precision choices matter too. The first 20 layers were trained and deployed natively in 4-bit NVFP4, with the final four layers in 8-bit MXFP8. Since the weights ship in the exact numeric format used during training, you skip the accuracy hit that typically comes with post-training quantization. That's a meaningfully different approach from just taking a full-precision model and compressing it after the fact.

Looping internally instead of writing out chains of thought raises an obvious safety question: if the model reasons more without showing its work, does it get harder to audit? Thinking Machines addressed this head-on, releasing a companion variable-loop research checkpoint (TML-d24-varloop-research, trained for one to five passes) and partnering with the National Deep Inference Fabric and Goodfire on mechanistic interpretability work. Early monitorability tests found looping extended the model's no-CoT reasoning horizon from 16 seconds at one pass to 37 seconds at three passes, cutting chain-of-thought length by more than half. Notably, the remaining reasoning traces stayed human-readable, and the model showed no uptick in covert sabotage on the LinuxArena control benchmark.
On the numbers, Inkling 2 Mini leads all sub-100B open-weights models on DeepSWE v1.1 (48.2%), SWE-bench Pro v2 Hard (77.9%), and Vals Index v2.2 (46.5%), plus 79.2% on SWE-bench Verified and 85.3% on DeepSearchQA. On Stanford's intelligence-per-joule benchmark running on an NVIDIA DGX Spark, it scored 0.703, well clear of Qwen3.6-35B-A3B (0.430) and Gemma-4-26B-A4B (0.159).
It's not a free lunch, though. Factual recall scales with total parameter count, not clever looping, so Inkling 2 Mini posts -8.6 on AA-Omniscience v1.1 with a 63.4% non-hallucination rate, trailing Qwen3.8-27B's -6.4 and 71.4%. It also lags on multi-turn chat (50.0% on MultiChallenge) and multi-constraint instructions (33.4% on ComplexConstraints v1.1), and testers have noted occasional repetitive debugging loops and clipped phrasing in non-English output.
Against Qwen3.8-27B specifically, the two models tie at 20.1 on the Artificial Analysis Intelligence Index. But Artificial Analysis reports Inkling 2 Mini uses about 2.8 times less compute, runs four times faster (9.0 minutes per task versus 38.2 minutes), and costs six times less per task ($0.59 versus $3.53 at first-party rates). It also beats Gemma 4 31B and Meta's Muse Glimmer on agentic software engineering and offers a much longer context window: 1 million tokens versus 262K and 131K respectively.
The licensing is straightforward too. Apache 2.0 lets anyone, startup or Fortune 500, download, modify, fine-tune, and embed the model in commercial products royalty-free, with no requirement to open-source what they build on top. That puts Inkling 2 Mini on the same legal footing as Qwen3.8-27B, Gemma 4 31B, and Muse Glimmer, and avoids the custom commercial terms attached to some larger frontier releases, like Mistral's Ecosystem Open-Weight license or Meta's Muse Model License for its 816B Muse Spark Open.
The developer reaction has been fast and largely positive. Unsloth AI shipped quantized GGUF builds with llama.cpp support within a day, touting a version that runs on 16GB RAM. Apple silicon specialist Exo Labs posted throughput numbers from its collaboration with Thinking Machines: 210 tokens/sec decode and 3,620 tokens/sec prefill on a 96GB M5 Ultra Mac Studio, 125 tokens/sec decode on a 48GB M5 Max. They also claimed the model built a native weather app 1.8x faster and 11x cheaper than Claude Sonnet 5.5. Independent developer Kyle Hessling got the model to autonomously generate a motion graphic about its own specs at 104 tokens/sec on an RTX 5090, calling it "the most impressive small model I've tested from anyone outside of Qwen." Some testers pushed back a bit, flagging occasional failures in game-mechanics generation and early quantization quirks in the still-maturing looped-model runtime ecosystem.
This is the fourth open-weights model Thinking Machines has shipped since founding in early 2025, following the 975B original Inkling, the 276B Inkling-Small, and the 594B Inkling 2 Small. The pace has been relentless, and it tracks a clear thesis: Mira Murati's team isn't just chasing frontier scale, it's betting that smarter compute allocation within a model can rival brute-force parameter growth. Inkling 2 Mini is the clearest proof point yet. For developers who want a private, high-agency coding assistant running entirely on a laptop or single workstation, without surrendering compute to the cloud, looped transformers just went from an academic curiosity to something you can download and run tonight.
Tags
Original Sources
Thinking Machines wows with Inkling 2 Mini looped transformer model that runs locally on laptops
↗ https://venturebeat.com/technology/thinking-machines-wows-with-inkling-2-mini-looped-transformer-model-that-runs-locally-on-laptops
About the author
Kai built ML infrastructure at a Bay Area startup before developing an obsession with transformer architectures and inference optimisation that eventually pulled him out of product work entirely. A stint at a compute research lab sharpened his instinct for what actually matters in a model release versus what is marketing. He writes from the inside — from the perspective of someone who has debugged the systems he is describing at three in the morning. He is allergic to hype and instinctively drawn to the unglamorous plumbing questions that everyone else skips over.
More from The Engineer →This Week's Edition
22 November 2026
10 articles
Related Articles

The Rise of Texting-Native AI Agents: A Field Guide
Products & Applications · 5 min

Microdrama Startup Shortical Tests Face-Licensing Deals as AI Replaces Live-Action Sets
Products & Applications · 5 min

Signal Blocks Screenshots on Mobile to Fight Back Against AI Agents Reading Your Chats
Security & Risk · 5 min
Related Articles

The Rise of Texting-Native AI Agents: A Field Guide
Products & Applications · 5 min

Microdrama Startup Shortical Tests Face-Licensing Deals as AI Replaces Live-Action Sets
Products & Applications · 5 min

Signal Blocks Screenshots on Mobile to Fight Back Against AI Agents Reading Your Chats
Security & Risk · 5 min
More Stories
© 2026 Cedar & Bloom. All rights reserved.