Speech models and a small omni quant lead the Hub today. NVIDIA added Nemotron 3.5 Transcribe to the blog and the model rose to the default page. ggml-org published a 1 GiB LFM3-Omni quant for local run. MiniCPM-OCR-1.0 entered the default 30 overnight. Papers rotated to a new daily slice with DTCA leading upvotes and a Falcon checkpoint with weights.
On the Hub
NVIDIA Nemotron 3.5 Transcribe 34B / 4.3B active OpenMDW-1.1
NVIDIA posted about Nemotron 3.5 Transcribe on 2026-10-28.1 The card entered the default trending page at #15. The model is 34B total with 4.3B active and uses OpenMDW-1.1.2 The blog claims 5.33% average WER on English. The README table from October 12 lists 4.71% mean WER on nine public sets. The BF16 file is 62.861 GiB. The quick start calls for one H100 80 GB. Quantized versions for RTX Spark are listed as coming soon.
ggml-org LFM3-Omni-2B-GGUF
ggml-org published LFM3-Omni-2B-GGUF on 2026-10-28. The Q4_K_M file is 1.024 GiB. The parent LiquidAI/LFM3-Omni-2B is default #25.3 The repo uses LFM Open License v1.0. It needs a llama.cpp build with LFM3-Omni support. The mmproj holds vision and audio encoders.
openbmb MiniCPM-OCR-1.0 Apache 1.3B
openbmb/MiniCPM-OCR-1.0 is default #24 this morning. Last modified is 2026-10-29T07:26:30Z. The card is Apache 2.0 and 1.3B.4 The README claims 93.68 on Real5-OmniDocBench. The card entered the default 30 overnight.
Daily Papers Oct 29 slice DTCA and JusperLee Falcon
Six of fifty daily paper wrappers are dated 2026-10-29. ZJU DTCA leads upvotes with 256 and has zero linked models.5 Tsinghua Falcon is on the same slice with 142 upvotes. The weights are JusperLee/Falcon at 45.1 MiB Apache, last modified 2026-10-29T01:05:36Z.6 The checkpoint is distinct from TII Falcon.7
