← All models

LFM2.5-24B-A2B-Base

Liquid AI · 2026-10-13 · 24B parameters

LFM2.5-24B-A2B-Base is the pre-trained text-only checkpoint from which Liquid AI builds LFM2.5-24B-A2B, released with open weights alongside it on October 13, 2026 and recommended for heavy fine-tuning. Liquid reports about 33T pre-training tokens followed by 2T mid-training tokens at 32K context and 300B long-context tokens at 131K context, and describes a 36-trillion-token training budget. It shares the 40-layer hybrid convolution and GQA MoE architecture (64 experts, top-4 routing, 2.45B reported active parameters) and 131,072-token context; the BF16 checkpoint contains 23,971,587,712 parameters, and the tokenizer assigns 125,017 distinct IDs within 128,000 embedding rows.

Benchmark scores

No benchmark scores recorded.