SHELFCRITTER AI Digest

COMPANY

Liquid AI

companytopic-noteopen-sourcefoundation-models

Overview

Liquid AI is a non-OpenAI foundation-model lab building efficient sparse-MoE small models for on-device and cost-sensitive inference, where the active-parameter envelope is the binding constraint.

Timeline

  • 2026-05-30-AI-Digest — Liquid AI announces LFM2.5, an 8B-parameter mixture-of-experts with 1B active params, trained on 38T tokens (Liquid AI blog). Reached HN front page at 170 pts / 62 comments. Why it matters: another efficient sparse-MoE small model from a non-OpenAI lab — relevant for on-device and cost-sensitive inference where the active-parameter envelope is the binding constraint.

  • 2026-10-15-AI-Digest — Liquid AI releases LFM2.5-24B-A2B, a 24B total / ~2B-active MoE positioned for low-latency on-device agent workloads (Liquid AI blog). Lands on HN at 125 pts / 34 cmts. Scales up from the May 2026 8B-A1B cut while holding the same active-parameter envelope that defines the lab’s thesis — the sweet spot for on-device agent loops is now a tight band around 20–35B total, 2–3B active, with DeepSeek-V4-Flash, MiMo V2.5 Pro, and Qwen3.6-35B-A3B in the same frame. Reframe worth carrying: another serious push into the small-active-params regime that makes agentic loops viable on laptops and phones, not scale-up abandons the LFM2.5 small-model lane. Log against MOC - Open Source Models.

Key Developments

  1. LFM2.5 — 8B-A1B MoE on 38T Tokens (May 30, 2026): 8B total / 1B active mixture-of-experts trained on 38T tokens, positioned for on-device and cost-sensitive inference. The active-parameter envelope (1B) is the binding constraint for the target deployment surface; the 38T-token training run is the magnitude signal (2026-05-30-AI-Digest).

See also: LFM2.5, MOC - Open Source Models.