Liquid AI releases LFM2.5-8B-A1B edge model

Liquid AI announced LFM2.5-8B-A1B, the follow-up to its October 2025 LFM2-8B-A1B. The company says the model "fits comfortably even on an entry-level laptop." The base and post-trained weights are available on Hugging Face and in Liquid's Playground.

The architecture still combines MoE, grouped-query attention and gated short-convolution blocks. Most of the other details changed:

• Pretraining grew from 12T to 38T tokens, followed by large-scale reinforcement learning.
• The context window went from 32,768 to 128,000 tokens.
• The vocabulary doubled from 65,536 to 128,000 tokens. Liquid extended the existing tokenizer rather than retraining from scratch. Text per token improved in all 16 languages tested, with the biggest gains in Thai (+238%), Hindi (+120%) and Vietnamese (+118%).
• It is now a reasoning-only model. Liquid argues that the small active parameter count makes each chain-of-thought token cheap.

Liquid also added a preference-optimization stage and an RL shaping reward to cut "doom loops" in long reasoning traces, such as repeated "Wait…" restarts. A separate RL stage rewards the model for declining to answer questions beyond its knowledge.

Against the earlier model, Liquid reports:

• AA-Omniscience non-hallucination rate: 7.46 to 63.47
• IFBench: 26.00 to 56.47
• BFCLv4: 25.52 to 48.50
• Tau² Telecom: 13.60 to 88.07
• AIME25: 20.00 to 42.53

Knowledge accuracy barely moved (7.33 to 8.67), so the hallucination gain comes mainly from the model declining to answer more often, not from knowing more.

Liquid says the model is competitive with dense models of similar total size and with much larger MoEs, and lists IBM's Granite-4.0-H-Tiny among the comparisons. The comparison table in the source is cut off, so the full set of competitor results is not available. All figures are Liquid's own; no independent evaluation has been published yet.