Thinking Machines’ open 26.4B Inkling 2 Mini loops 12 of its 24 layers three times per token and fits in about 20GB. Most scores are self-reported.
Tag: on-device AI
A Hugging Face post runs a 180B open MoE model on a 32GB MacBook by streaming experts from the SSD. It’s about 6% behind a rival engine.
llama.cpp, the engine behind much local AI, reached 1.0. The real news isn’t speed but a promise: its API won’t break without warning.
Zyphra’s open ZONOS3-Mini clones a voice from a clip of up to 60 seconds, in 43 languages, on a phone. Its docs say nothing about consent.
Tether’s open 0.8B BotPsy calls tools on-device and doubles its base model’s agent scores. Its makers say: keep it from money. Scores are self-reported.
Liquid AI’s open-weight LFM3: six ~2B models for text, vision, audio and decisions, grown from one base. The scores are self-reported.
Recursive’s open-weight Flint codes on a 48GB laptop, and its makers say an AI research system helped design it. The scores are self-reported.
Nvidia’s open Nemotron 3.5 Bolt squeezes a 120B parent into 74B parameters and 24GB, with 2-bit experts. What the model card shows, and what it skips.
Liquid AI’s open-weight LFM2.5-24B-A2B runs agents on laptops with 2B active parameters. Strong at tools, weaker on facts, by its own admission.
Liquid AI’s open-weight d1 models answer yes/no, pick-one and score questions about text, images and speech in one pass, without writing text.