Google listed a Gemma 4 speaker-diarization post-processor on the Hub on 4 October, and the official 5.3 GB GGUF is already in the repo. A community blog dated 5 October walks through DeskMind, a Mac agent that will not click until a 0.8B Qwen answers a multiple-choice question. ggml-org put its own Q4_K and Q2_K of GLM-5.3-Flash on the Hub on 3 October, with the vision projector and speculative-decoding sidecars in the same repo.
On the Hub
Google DiarizationLM: Gemma 4 cleans who said what
Google listed DiarizationLM-Gemma-4-E4B-v1 on the Hub on 4 October. The card is a speaker-diarization post-processor on Gemma 4 E4B (card: 4B dense), Apache 2.0, with an official Q4_K_M GGUF of about 5.30 GB in the same repo.1 It is an LLM cleanup pass after ASR, not an acoustic frontend, and not last month's NVIDIA diarization model. Training used Locality-Preserving Oracle Supervision on Fisher, Callhome, ICSI, and AMI, rather than two-speaker telephone data only.
The card claims Fisher WDER of 5.32 percent on the USM plus turn-to-diarize baseline falls to 2.99 percent, and Callhome WDER of 7.74 percent falls to 4.92 percent. The older Llama 3 8B DiarizationLM sits between those numbers on the telephone sets, at twice the named parameter count. The card banner says this is not an officially supported Google product.
DeskMind: a 0.8B Qwen that asks before it clicks
gxcsoccer published a Hugging Face community blog on 5 October about DeskMind, an open-source macOS computer-use agent whose every step is a typed multiple-choice decision. Two Qwen3.5 fine-tunes run locally on MLX: brain-0.8b answers first, and brain-4b retries below 0.96 confidence or on a risky step. eyes-4b finds the click target when an app has no accessibility tree.2 When two rows both look like Lisa Wong, the agent is trained to ask instead of guessing a click.
The models have been on the Hub since 27 September. The 5 October write-up is the harness, not a same-day weight drop. The server speaks POST /v1/systemone, the same request shape llama.cpp already uses for decision models. The release revision is 8-bit MLX under Apache 2.0.
The author reports 39 of 39 correct on a 13-task desktop bench run three times on one M4 Pro, with zero false "done." The 4B scores 0.835 on public JevBench v1.4.2, with a sealed run still pending. About 70 percent of steps still escalate to the 4B. Those scores are the author's own runs, with the limits attached.
ggml-org GLM-5.3-Flash-GGUF: official Q4_K of a 320B MoE
ggml-org auto-converted GLM-5.3-Flash, a 320B MoE with 18B active, into llama.cpp GGUF on 3 October. The Q4_K pack is about 181 GB and the Q2_K pack is about 134 GB, with a Q8_0 vision mmproj and MTP sidecars for speculative decoding in the same repo.3
The Hub widget's 4.95 GB 4-bit file is the MTP sidecar, not the model. Low-bit quants have no imatrix calibration. This is the official ggml-org convert, a different pack from Unsloth's August GGUF of the same base, and it wants a couple of hundred gigabytes of disk rather than a 5090.
