LFM2.5-24B-A2B

Released 2026-10-12 on Hugging Face (API createdAt; card updated 2026-10-13). Affordable-inference class: 24B total MoE with 2.45B active — under the sub-8B-active bar, and smaller-active than the 35B-A3B reference class. Vendor card: 24.0B total, 2.45B active (the name says A2B), 40 layers of which 30 are double-gated short convolutions and 10 are GQA, 64 experts top-4, vocab 128000 (doubled from LFM2-24B-A2B's 65K via in-place tokenizer expansion), context grown from 32K to 131072. config.json agrees: hidden 2048, heads 32/8 so head_dim = 2048/32 = 64, gqa_ratio 4, intermediate_size 11776 on the 2 dense layers (num_dense_layers 2), moe_intermediate_size 1536, num_experts_per_tok 4, no shared expert, rope_theta 1e7, tie_word_embeddings true. KV (bf16, attention layers only): 10 full_attention layers x 8 kv heads x 64 head_dim x 2 x 2 bytes = 20480 B/token = 20 KiB/token. flops_per_token_active = 2 x 2.45e9 = 4.9e9, the schema approximation; hybrid conv blocks mean the real matmul count is not exactly 2*active_params. THROUGHPUT (vendor_claim, card): 184 tok/s on an Apple M5 Max, and up to 21K aggregate output tok/s on one NVIDIA B300 (SGLang 0.5.22, BF16, 1024 in / 256 out, concurrency sweep). Card also claims ~1.7x faster than comparable MoEs on-device and LiveClaw 48.09 vs Qwen3.6-35B-A3B 38.19 — vendor numbers, not independently measured here. Text-only; always-on <think> in the chat template. Base checkpoint is LiquidAI/LFM2.5-24B-A2B-Base (created 2026-10-09, same shape). License is the LFM Open License, not Apache.