Liquid posted a team article on October 13 for LFM2.5-24B-A2B, a 24B on-device agent with 2.45B active and an official 13.53 GB Q4_K_M GGUF. NVIDIA published a coding-agent skill the same day that domain-adapts Nemotron 3.5 ASR, with an author-run air-traffic-control cut from 77 percent word error to about 22. Korea University's DMIS Lab wrote up UltraBERT, a pair of Apache English encoders built for both classification and retrieval. FINAL-Bench put VKORE-46M-NPU on the Hub this morning as a TinyStories pretrain on inference-only NPUs.
On the Hub
LiquidAI LFM2.5-24B-A2B
Liquid published a Hub team article on October 13 for LFM2.5-24B-A2B, a 24.0B MoE with 2.45B active, 131k context, and reasoning before every answer. The instruct weights have been on the Hub since October 12. The official GGUF Q4_K_M is 13.53 GB, last updated October 13, with an Unsloth Dynamic GGUF the same afternoon. Liquid's own table scores 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B and claims 184 tok/s on an Apple M5 Max, author numbers rather than an independent eval. The license is LFM Open License v1.0, which blocks commercial use above $10 million in annual revenue.12
NVIDIA Speech Agent Skills
NVIDIA published a Hub post on October 13 for Speech Agent Skills, a coding-agent loop that domain-adapts Nemotron 3.5 ASR through the nemotron-asr-finetune orchestrator plus a NeMo specialist. The base streaming 0.6B checkpoint has been on the Hub since May 15. In NVIDIA's own air-traffic-control study, staged adaptation cut word error from 77.00 percent to 22.46 percent on a two-hour human-annotated ATCO2 set. General-English word error moved from 3.50 percent to 3.63 percent. Those figures are author-run on NVIDIA's protocol.34
UltraBERT 150M / 380M
Korea University's DMIS Lab published a Hub post on October 14 for UltraBERT, a pair of Apache-2.0 English bidirectional encoders at 150M and 380M with native 32K context. The weights have been on the Hub since October 10, so the news is the write-up. The author's 380M table is 91.0 on GLUE and 47.7 on BEIR, and UltraBERT-Embed-150M is 67.24 on MTEB (eng, v2). Those scores are author-claimed, and loading needs transformers 5 with trust_remote_code.56
VKORE-46M-NPU
FINAL-Bench published a Hub post this morning for VKORE-46M-NPU, a 45.8M decoder-only TinyStories model pretrained from scratch on 32 Rebellions ATOM-Max inference NPUs. The weights landed on the Hub the same morning. The author's card says training loss fell from 5.65 to 0.42. The kernels are withheld as patent-pending, so this is a proof of concept, not a local assistant.78
