Skip to content
LIVE
The Executives BriefThe Executives Brief

Murati's Inkling 2 Mini ties Alibaba's 27B flagship while running on a MacBook

A 26.4B open-weights model that runs offline on a laptop now matches Alibaba's flagship on intelligence while costing six times less per task, and that changes the math on every agentic coding budget.

ByOmar Al-BalawiTechnology Correspondent, The Executives Brief
·4 min read
Murati's Inkling 2 Mini ties Alibaba's 27B flagship while running on a MacBook
Executive summary

Thinking Machines Lab, the San Francisco startup led by former OpenAI CTO Mira Murati, released Inkling 2 Mini, a 26.4B-parameter open-weights model that runs offline on a 32GB MacBook Pro and ties Alibaba's dense 27B flagship on aggregate intelligence at roughly six times less cost per task. For decision-makers, it reframes local deployment as a cost and data-residency strategy rather than a capability compromise, while exposing factual recall as the trade-off.

Mira Murati's Thinking Machines Lab released Inkling 2 Mini yesterday, and the number that should make every infrastructure budget owner sit up is this: a 26.4 billion parameter multimodal model that runs offline on a 32GB MacBook Pro or a single desktop GPU, ships under an Apache 2.0 license, and ties Alibaba's dense 27B flagship on aggregate intelligence. Both score 20.1 on the Artificial Analysis Intelligence Index v5.0. The difference is what it costs to get there. Artificial Analysis reports Inkling 2 Mini uses roughly 2.8 times less compute, runs four times faster at 9.0 minutes per task versus 38.2 minutes, and costs six times less per task, $0.59 against $3.53 at first-party API rates.

Standard transformers process each token in one linear pass, first layer to last. When a problem needs more sequential reasoning than the model's depth allows, the model has to think out loud, generating long chains of thought in text. That burns output tokens, adds latency, and inflates the bill. Inkling 2 Mini takes a different route. Per the model card, it is a 24-layer, decoder-only sparse Mixture-of-Experts system that routes each token to 8 of 128 experts, activating 3.4 billion parameters on a standard pass. The twist is how those 24 layers execute: the first six, the prelude, and the final six run once, while the middle 12 run three times using shared weights. That yields 48 effective layer applications and 5.4 billion effective active parameters per token without growing the storage footprint. Attention residuals let each pass read a gated message from the previous pass's middle layers alongside the initial prelude output. 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.

Getting that architecture onto consumer silicon took aggressive surgery. Thinking Machines pruned two-thirds of its experts mid-training, from 384 down to 128, then healed the 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 just as much: the first 20 layers were trained and deployed natively in 4-bit NVFP4, the final four in 8-bit MXFP8. Because the weights ship in the exact numeric format used during training, developers sidestep the accuracy degradation that typically plagues post-training quantization. That is the quiet engineering decision that lets a roughly 20GB download behave like a model twice its size.

On standardized evaluations it punches well above its weight class in exactly the areas enterprises are buying for. It leads all sub-100B open-weights models on DeepSWE v1.1 at 48.2%, SWE-bench Pro v2 Hard at 77.9%, and Vals Index v2.2 at 46.5%, while scoring 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, comfortably ahead of Qwen3.6-35B-A3B at 0.430 and Gemma-4-26B-A4B at 0.159. Against Gemma 4 31B and Meta's Muse Glimmer, the lead on agentic software engineering and long-context retrieval is commanding: 1 million tokens of context versus 262K for Gemma 4 31B and 131K for Muse Glimmer.

Looping cannot conjure facts out of thin air. Factual recall scales primarily with total parameter count, and that is where the compact design shows its seams. Inkling 2 Mini scores -8.6 on AA-Omniscience v1.1 with a 63.4% non-hallucination rate, trailing denser peers like Qwen3.8-27B at -6.4 and 71.4%. Thinking Machines' own documentation notes the model trails Qwen3.8-27B on multi-turn chat, 50.0% on MultiChallenge, and multi-constraint instructions, 33.4% on ComplexConstraints v1.1, and can occasionally fall into repetitive debugging loops or produce clipped phrasing in non-English languages such as Chinese. That defines the deployment boundary clearly: agentic coding and tool use, not encyclopedic recall.

Looping also raises a governance question boards should flag now. If a model does more reasoning internally without writing it down, does it become harder to oversee? Thinking Machines shipped a companion variable-loop checkpoint, TML-d24-varloop-research, trained for one to five passes, and partnered with the National Deep Inference Fabric (NDIF) and Goodfire to study mechanistic interpretability. In preliminary monitorability tests, looping extended the model's no-chain-of-thought reasoning horizon from 16 seconds at one pass to 37 seconds at three passes and cut chain-of-thought length by more than half, yet the remaining reasoning traces stayed human-legible, and the model showed no upward trend in covert sabotage on the LinuxArena control benchmark. That is a preliminary result, not a clean bill of health.

The demos are the part that travels. Thinking Machines ran the model in airplane mode on a 48GB M5 Pro MacBook Pro to build a playable flight simulator in minutes, and powered 20 persistent autonomous villagers simultaneously in a local town simulation. It is natively multimodal, accepting UTF-8 text and pixel-based images optimally between 40px and 4096px across a 1 million token context window, and outputting text. Weights are on Hugging Face under Apache 2.0 with an acceptable use policy. Hosted pricing runs $0.15 input and $0.50 output per million tokens first-party, from $0.08 and $0.35 on OpenRouter. Local runtimes including Ollama and EXO already support it, with 16GB+ machines served via GGUF.

The strategic read is that the open-weights frontier is splitting into two races. One is raw capability, where parameter count still buys factual recall. The other is intelligence per joule and per dollar, where architecture, not scale, decides the winner. Inkling 2 Mini is a credible shot at the second race, and it arrives with a U.S.-built, Apache 2.0 license at a moment when enterprises are under pressure to keep sensitive workloads on hardware they control. For CFOs, the six-times-cheaper-per-task figure is the one that moves budget lines. For security leads, offline execution on a laptop is a data residency answer. For every team currently paying frontier API rates to run agentic coding loops, the question just shifted from whether a local model is good enough to which tasks still justify the cloud.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Technology