Day 7 receipt
LFM3
Day 2 of 7Liquid AI's open-weight 2B-class multimodal model family, built on a new hybrid architecture and aimed at on-device and fine-tuned use.
Launch dayWhat they claimed, quoted word for word · 3
“a new tri-hybrid architecture and training recipe allow LFM3-1.7B to outperform our previous LFM2.5-2.6B while using 35% fewer parameters.”
No independent side-by-side with LFM2.5-2.6B yet; one user saw a wrong text answer.
“the text model fits in about 1 GB at 4-bit precision.”
One user could not load the MLX builds; no memory measurement reported yet.
“256K context window for the text model.”
No user has reported long-context tests yet.
Day 7What real users report · 2
A user tried the omni and text models as MLX builds in several Mac runners and none would load, with a safetensors read error.
Comment in a 170-point launch thread (as of 29 Oct) · Open the source
A user said the vision-language model gave better text-only answers than the 1.7B text model in the playground, which answered a simple animal-sound question wrongly.
Comment in a 170-point launch thread (as of 29 Oct) · Open the source