LiquidAI/LFM2.5-24B-A2B model card and config.json

Primary source for the LFM2.5-24B-A2B record. Model card (fetched 2026-10-16) states 24.0B total / 2.45B active, 40 layers (30 double-gated short convolution + 10 GQA), 64 experts per MoE block with top-4 routing, vocab 128000, context 131072. HF API createdAt 2026-10-12. config.json (resolve/main, same fetch) confirms model_type lfm2_moe, hidden_size 2048, num_hidden_layers 40, num_attention_heads 32, num_key_value_heads 8, intermediate_size 11776, moe_intermediate_size 1536, num_experts 64, num_experts_per_tok 4, num_dense_layers 2, max_position_embeddings 131072, vocab_size 128000, rope_theta 10000000, tie_word_embeddings true, and layer_types = 30 conv + 10 full_attention. License cardData is 'other' (LFM Open License). Throughput numbers on the card (184 tok/s on an Apple M5 Max; up to 21K aggregate output tok/s on one B300 with SGLang 0.5.22, 1024 in / 256 out, BF16) are vendor claims. Architecture paper cited by the card is arXiv 2511.23404 (LFM2 technical report, Nov 2025), which covers the LFM2 family rather than this 24B-A2B release.