SHELFCRITTER AI Digest

MODEL

LFM2.5

modeltopic-noteopen-sourcemoeliquid-ai

Overview

LFM2.5 is Liquid AI‘s 8B-parameter mixture-of-experts foundation model with 1B active parameters, trained on 38T tokens. Positioned for on-device and cost-sensitive inference where the active-parameter envelope is the binding constraint rather than raw parameter count.

Timeline

  • 2026-05-30-AI-Digest — Liquid AI announces LFM2.5 as an 8B-A1B MoE trained on 38T tokens (Liquid AI blog). HN front-page coverage at 170 pts / 62 comments. Continues the trajectory of efficient sparse-MoE small models from non-OpenAI labs targeting on-device and cost-sensitive inference.

  • 2026-10-15-AI-Digest — Liquid AI ships LFM2.5-24B-A2B, a 24B total / ~2B-active MoE positioned for low-latency on-device agent workloads (Liquid AI blog). HN surface at 125 pts / 34 cmts. The variant scales the family up roughly 3× in total parameters while holding the active-parameter envelope at ~2B — practitioner-relevant for the agentic-loop deployment surface where per-token compute dominates. Positioned alongside DeepSeek-V4-Flash and MiMo V2.5 Pro as another serious push into the small-active-params regime that makes agentic loops viable on laptops and phones. Log against MOC - Open Source Models.

Key Developments

  1. 8B Total / 1B Active MoE Architecture: Sparse-MoE shape with a 1B active-parameter envelope keeps per-token compute low while preserving the representational capacity of an 8B-total model. The active-parameter envelope is the practitioner-relevant figure for the target deployment surface (2026-05-30-AI-Digest).

  2. 38T-Token Training Run: At the higher end of public training-token counts for sub-10B-active models, indicating Liquid AI is investing in data scale even at the small-model tier rather than under-training.

See also: Liquid AI, MOC - Open Source Models.