No Hype: Liquid AI’s new model is fast, but slightly dishonest
Efficiency gains and massive subsidies drive the industry. We look at Liquid AI’s lightweight model, Nvidia’s European expansion, and DeepSeek’s new toolkit.
Liquid AI’s small model is surprisingly fast
Illustration generated with AI on our own machine.
What happened. Liquid AI released the LFM2.5-24B-A2B model with 24.0B total parameters and 2.45B active parameters, according to unite.ai on 14 October 2026. The model features a 131,072-token context window and a 128,000-token vocabulary (unite.ai). Pipette benchmarks measured 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395.
What we think. Benchmarks are easy to game. We prefer seeing how a model handles a real workload.
The good bit. Running capable models locally reduces cloud dependency and improves privacy.
What you can do. Check your available memory using our VRAM calculator before downloading large weights.
Nvidia’s European expansion may stifle local competition
Illustration generated with AI on our own machine.
What happened. Nvidia plans to announce $2 billion in European investments this week, according to ainvest.com on 14 October 2026. The European Commission is offering €10 billion in public financing for seven AI gigafactories to attract €20 billion in private investment (ainvest.com). The Commission’s report notes that Europe’s top five cloud service providers are all American. Nvidia has already placed 35 new AI supercomputers across 23 European countries.
What we think. We are essentially paying American chipmakers to build their own fortress in our backyard.
The good bit. The massive influx of capital and infrastructure could accelerate the deployment of AI services across the continent.
What you can do. Stay informed about upcoming regulations by reading our EU AI Act guide.
DeepSeek Harness brings plugins to the desktop
Illustration generated with AI on our own machine.
What happened. The DeepSeek team released the DeepSeek Harness desktop app for macOS and Windows on 30 September 2026, according to incrypted.com on 14 October 2026. The V4.1-Flash model, released in September 2026, supports context windows of up to 1 million tokens (incrypted.com). The first version of the software was published on GitHub on 13 August 2026.
What we think. Plugins are just a way to make software convenience look like architectural progress.
The good bit. The availability of a desktop app simplifies the process of managing autonomous agent workflows.
What you can do. Explore new ways to build local AI for developers using these new tools.
In one line
Local efficiency, massive subsidies, and new developer tools define this week’s shifts.
Work with us
If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.