Brief IA
›

Inkling 2 Mini Works on a Laptop

Creative AI·Tom Levy·

Inkling 2 Mini Works on a Laptop

Inkling 2 Mini Works on a Laptop
Key Takeaways
1Thinking Machines unveiled Inkling 2 Mini on Friday, an open-weight model for laptops.
2Its use with Ollama on Mac requires at least 32 GB of memory.
3Artificial Analysis ranks it 20th in its Intelligence Index.
Why it matters : the model can operate offline on a laptop, while Thinking Machines is studying the monitoring of its internal reasoning.

Thinking Machines Lab, the startup founded by Mira Murati, unveiled Inkling 2 Mini on Friday, an open-weight model capable of running on a laptop. Its architecture repeats internal calculations rather than explicitly detailing its reasoning, a property the lab is also studying for monitoring purposes.

Internal reasoning remains difficult to monitor

To control their models, labs examine the chain of thought, that is, the written reasoning. However, loop architectures can perform more reasoning internally without exposing it to a monitoring device. OpenAI expressed this concern in September, following Astra's strong performance in tests. Thinking Machines also acknowledges the importance of this unwritten reasoning for monitoring.

The lab offers a research version with a variable number of loops, ranging from one to five, intended for third-party experimentation. In its trials, five loops reduced the written reasoning to 1,661 tokens, compared to 3,703 with one loop, which is less than half. Thinking Machines indicates that its readability remained the same and that the additional loops did not facilitate the concealment of sabotage.

The lab does not observe, at this scale, any significant degradation in monitoring reasoning but qualifies its conclusions as preliminary. The experimental model received reinforcement learning only with three loops; more extensive training could enhance the compression of reasoning. A device limited to observing tool calls and outputs also missed a significant number of attacks during testing.

Use on Mac and offline demonstrations

Available on Hugging Face under the Apache 2.0 license, the weights can be used with Ollama on Macs equipped with at least 32 GB of memory. The model occupies about 20 GB and accepts a context of 1 million tokens.

To demonstrate its local functionality, researchers from Thinking Machines conducted a demonstration during a flight: on a disconnected MacBook Pro in airplane mode, the model produced a playable flight simulator in just a few minutes. Another demonstration showcased 20 pixel villagers from a town it had created, all running on a single MacBook without a connection.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

This version follows the first Inkling, launched in July, a very large open model that Thinking Machines acknowledged was not the best. In the realm of local execution, Flint, developed by Recursive, is among the new competitors. It is also included in the lab's comparative programming charts.

Programming tests and task costs

The independent organization Artificial Analysis ranks Inkling 2 Mini 20th in its Intelligence Index, on par with Qwen3.8 27B, with about five times fewer active parameters. Its measurements indicate a cost of $0.59 and a duration of nine minutes per task, compared to $3.53 and 38.2 minutes for Qwen3.8 27B. It describes the model as approximately six times cheaper and four times faster per task. At equal scores, Inkling produces about 38% more tokens in output; in these measurements, OpenAI's GPT-6 Luna scores 28 at $0.15 per task.

Thinking Machines claims an advantage over all open-weight models under 100 billion parameters in programming tests such as DeepSWE. The lab reports 48.2% on DeepSWE v1.1 and 77.9% on a challenging selection from SWE-bench Pro. In its comparison, Alibaba's Qwen3.8-27B achieves 42.2% and 73.8%, respectively.

However, the lab acknowledges Qwen3.8-27B's advantage in multi-turn exchanges and instructions that accumulate numerous constraints. Inkling 2 Mini also lags behind larger models in difficult reasoning tests. According to Thinking Machines, the loops do not enhance its ability to memorize facts.

Central blocks repeat their calculations

Inkling 2 Mini is based on a mixture of experts: it has 26.4 billion parameters, with 3.4 billion activated for each token. Its 24 layers form four blocks of six: the end blocks pass once, while the two central blocks are executed three times. Including loops, the total reaches 5.4 billion weight activations per token.

The majority of the weights were trained and then executed directly in Nvidia NVFP4 format, on 4 bits, without loss related to post-training conversion. The lab also removed one-third of the experts during training. It then retrained the model with 1 trillion tokens from a larger internal Inkling.

According to Thinking Machines' trials, removing the loops reduced training costs by 7.5%. However, this variant then required 140% more computation at runtime to achieve the performance of the looped version.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.