AI Momentum 98.5 +1.5%GitHub Pulse 5503 +6.5%Capital Flow $204.0B +19039.3%Hiring Heatmap 0.0 -100.0%
Advertisement
Open Source

Thinking Machines' new free AI model runs on a well-equipped laptop and ties Alibaba's Qwen on an independent test at a sixth of the cost

Thinking Machines Lab, the AI startup led by former OpenAI technology chief Mira Murati, released a model on Friday that anyone can download and run for free. Called Inkling 2 Mini, it is small enough to run on a well-equipped laptop.

For a business, that means AI work with no per-use cloud bill and no data leaving the building. The lab is also using a standard license with no size caps or permission rules, which makes legal sign-off simpler.

The strongest claim has outside support. Artificial Analysis, an independent firm that tests AI models, gave Inkling 2 Mini a score of 20 on its overall intelligence index. That ties Qwen3.8-27B, a popular free model from Alibaba.

Inkling 2 Mini got there for far less. Artificial Analysis put the cost of running its test suite at $0.59 per task, against $3.53 for the Qwen model. It also finished about four times faster, at 9 minutes per task versus 38.2. The cost figure uses Thinking Machines' own hosted prices.

What buyers get

The model has 26.4 billion parameters, a rough measure of its size. Only about 3.4 billion are used for each word it produces, which keeps it cheap and quick to run.

Its whole footprint is about 20 gigabytes of memory. The lab says it runs on any Mac with at least 32 gigabytes of memory through Ollama, a popular free app for running AI models locally. A compressed version listed on LM Studio, a similar app, needs at least 18 gigabytes.

The unusual part is the design. Thinking Machines calls it a "looped" model: it runs its middle layers three times for each word. The idea is that the model does more thinking internally instead of writing out long chains of reasoning.

The lab says a version without looping was slightly cheaper to train but needed 140% more computing power when in use to match it. That figure has not been independently checked. The model can also take in very long documents, up to about a million tokens, or units of text, in one request.

It is released under Apache 2.0, a widely used open-source license that allows commercial use, alongside an acceptable-use policy. That stands apart from the custom licenses some rivals attach to their free models, which can add revenue limits or require permission for big users.

Where the claims are thinner

Most of the headline coding results come from the company itself. Thinking Machines says Inkling 2 Mini leads all free models under 100 billion parameters on several coding tests. On one, DeepSWE, it reports 48.2% against 42.2% for the Qwen model.

The lab is open about weak spots. It says the model trails Qwen on multi-turn conversations and on instructions with many conditions. Looping, it says, does not help the model hold more facts.

That shows up in factual accuracy. On a test of how often a model avoids making things up, Inkling 2 Mini scored 63.4%, below the Qwen model's 71.4%. Artificial Analysis also found it produces about 38% more text to reach the same score, according to The Next Web.

The savings case also has limits. On the same Artificial Analysis index, OpenAI's paid GPT-6 Luna, at its highest setting, scored 28 at $0.15 per task. In other words, a closed model sold by OpenAI was both smarter and cheaper per task on this test. Inkling 2 Mini's edge is against other free models, and for firms that want to run AI on their own machines.

There is an oversight question too. Safety teams often watch a model's written reasoning to catch bad behavior. A model that thinks more internally gives them less to read. The lab says early tests showed looping did not meaningfully reduce how easy the model is to monitor, but calls the results preliminary. It also found that a monitor watching only the model's actions and outputs missed a significant number of test attacks.

Why it matters for investors

Thinking Machines has built its strategy around releasing models for free. Inkling 2 Mini follows Inkling 2 Small, a much larger model released earlier this month. The new model is also available on Tinker, the lab's hosted service for running and customizing models.

Investors are paying up for that approach. The lab raised $2 billion at a $12 billion valuation in July 2025. The Information reported in September, and other outlets in early November, that it was in talks for a round of roughly $5 billion or more at a $40 billion valuation before the new money, led by Accel. Those reports did not have the company confirming a deal.

For corporate buyers who want a US-built free model, one with a standard license and independent scores that match Qwen adds another option.

What to watch next: independent tests of the coding claims, download and usage numbers in the coming weeks, and whether looped designs prove as easy to oversee as the lab's early results suggest.

Written by AI from the public sources linked below.