HF breakouts Sep 21–28: twelve verified >10x jumps, led by a 340M classifier at 2,110x

Evidence cutoff: September 28, 2026, 09:00 PT. Twelve Hugging Face repositories cleared this week's screen, which asks for two dated readings of the same model's download counter, both inside the seven-day window, at least ten times apart. The widest was 2,110x, on a 340M classifier. Audio returned to the table with two entries after an empty week. Four of the twelve are derivatives of one image model released on September 14, and every one of those four carries the same non-commercial licence.

The counter is Downloads last month, the rolling 30-day figure printed on every model page. One reading says how much attention a repository has drawn over the past month, and direction appears only when two readings are compared. Both readings appear for every entry below, so the ratio can be recomputed rather than taken on faith. Unless a row says otherwise, both readings come from archived copies of the model page captured at the times shown, and where the second reading is the live Hub counter the row says "at cutoff".

One caveat changes how a few of these numbers should be read. A repository published during the window can print no monthly figure at all for its first days, so a ratio computed from its first available reading understates the move. That applies to MiMo-V2.6-Pro-RL, ThinkingCap-Qwen3.8-27B-GGUF and Pruna-Qwen-Image-2.1, and each of those rows carries the note.

All three LLM breakouts this week are small specialists, and all three are Apache 2.0. Each one does a narrow job in a single forward pass: given a state and a set of labels or typed questions, it returns a choice with a probability attached. None of them writes prose.

GLiNER2.5-Decide had the widest move of the week, from 7 to 14,772 in two days, and it is the smallest model in the table at 340M parameters. Fastino built it on a DeBERTa-v3-large encoder. The label set is an argument rather than a training run: pass intent, routing, sentiment, priority, policy, document type, moderation severity or spam categories at call time and the model returns a label, or every label above a threshold. There is no prompt template and no generated token. The card reports 60.2% exact-match accuracy across a 17-domain held-out suite, ahead of GLiNER2.5-Decide-1B at 59.6%, JevK5 at 57.6% and Laya Router at 46.6%, and it runs on CPU or GPU through the gliner2 package. 3

The first test takes an afternoon. Write down the labels your product already routes on, replay a few hundred logged requests, and count how often this model picks the label your production router picked. Adding a new category afterwards is a code change.

Kev-9B and Kev-4B come from the same design and the same author. Kev-9B is a LoRA with 45.4M trainable parameters plus a pointer head on Qwen/Qwen3.5-9B-Base, and it returns a probability distribution per typed question while serving a documented API contract. Its card reports 0.852 accuracy on a locked out-of-domain test the author says was read once, and the lowest Brier score of the Kev line at 0.237. Kev-4B is the smaller sibling, and its card logs a September 24 update that gained 26.3 points on the author's hard-v1 suite and 13.4 on devtools-v1, with the locked test unchanged. Apache 2.0 covers the adapter, the head and the Qwen3.5 base. 1733

Those accuracy figures are the author's own, measured on the author's suites. The number that decides anything is agreement with the routing or triage decision you already run, measured on your own logged traffic.

Four of the twelve entries are derivatives of Qwen-Image-2.1, the image model Qwen shipped on September 14 under the Qwen Research License. That agreement, dated September 20, grants rights to use, reproduce, distribute and modify the materials "FOR NON-COMMERCIAL PURPOSES ONLY" and sends commercial use to a separate licence from Qwen. 34 Every one of the four carries those terms, and the downloads went to the community builds that make the model loadable in ComfyUI.

The text encoder is the largest of the four, at 36.1x. pottokao repackaged Qwen3-VL-8B-Instruct, the model Qwen-Image-2.1 uses as its text encoder, in GGUF, FP8 and bf16 form with refusal-ablated encoder weights. The encoder is Apache 2.0 because Qwen ships Qwen3-VL-8B-Instruct under Apache 2.0, and the diffusion transformer it feeds stays under the research licence, which keeps the assembled pipeline non-commercial. 15

The card also ships a ComfyUI node that fixes two loading failures, and that is the reason this repository moved. The GGUF path was dropping the Qwen3-VL vision tower and returning hidden states of the wrong width, so reference images for editing never reached the vision model; the second fix handles DiT GGUFs that arrive without general.architecture metadata and were being rejected as an unknown architecture. Sizes run from 5.0 GB for the Q4_K_M encoder with a 1.2 GB projector, through 9.3 GB FP8, to 17.5 GB bf16. The card's sampling advice for the base model is 25 steps, with cfg 1.0 when the image holds no text and a higher value when it does. 15

0xSojalSec/Qwen-Image-2.1-Uncensored-GGUF holds five GGUF builds of the transformer, from a 4.05 GiB Q4_0 to a 7.59 GiB Q8_0. Its card repeats the abenzerps card with the file names changed, and the download links inside it still point at the abenzerps repository, whose counter read 1,062,921 at the cutoff. Start from the larger repository and treat this one as a mirror. 20

PrunaAI/Pruna-Qwen-Image-2.1 is the one derivative that changes the compute rather than the packaging. Pruna AI ships few-step LoRA adapters that generate in 5 or 8 steps with no classifier-free guidance, and the vendor measures them at up to 6.3x faster than the base pipeline. 25 e-n-v-y/Qwen-Image-2.1-Fix is a LoRA plus a workflow whose author says it repairs most of the base model's generation artefacts; the repository declares no licence, so the research terms travel with the weights. 27

Three of the week's largest absolute moves sit below the screen because their counters could not be paired. abenzerps/Qwen-Image-2.1-Uncensored-GGUF printed no monthly figure on September 20, the day it appeared, read 182,313 by September 22 and 1,062,921 at the cutoff, which is 5.3x between the two readable dates. 3536 unsloth/Qwen-Image-2.1-GGUF read 26,739 on September 23 and 194,341 at the cutoff, which is 7.3x. 3738 Viggle/Qwen-Image-2.1-viggle-turbo read 12,425 on September 24 and 101,512 on September 26, which is 8.2x. 3940 A ranking by absolute downloads puts those three first, and the screen ranks by ratio. All three are derivatives of the same non-commercial release.

The audio bucket was empty last week. This week it holds the second- and third-widest moves on the list, and both are speech models with terms a builder can work with.

NVIDIA's Nemotron-3-Diarization answers who spoke when. One checkpoint covers streaming and offline inference for up to eight speakers, and it orders its output channels by each speaker's first arrival so the labels hold across a long recording. The input buffer goes as low as 80 ms, the recommended low-latency setting is 0.32 s, the offline-style configuration uses a 30.4 s buffer, and the output frame resolution is configurable in multiples of 10 ms. Chunked inference carries no maximum audio length. NVIDIA released the model on September 23 and ships two runtimes: NeMo-Speech.cpp, where nemo-speech diarize meeting.wav returns the speaker turns locally, and NVIDIA NeMo Speech for training and fine-tuning. The same tool can attach word-level speaker tags to a transcript with nemo-speech transcribe meeting.wav --diarize --json. The licence is OpenMDW-1.1, and the card states the model is ready for commercial or non-commercial use. 6

Test it on the format your product actually handles. Meeting recordings, call-centre audio and interview footage each punish diarization differently, and a speaker who leaves the room and returns is the case that separates a usable model from a demo.

moondream/parakeet-redux is a 1.58-bit ternary build of nvidia/parakeet-tdt-0.6b-v3, in which every encoder weight is -1, 0 or +1. The whole model is 178 MB and transcribes 25 languages at 113x real time on eight x86 CPU cores, which the card says is 2.5x the fastest other Parakeet runtime it measured. Accuracy sits close to the original rather than on top of it: 6.55 average word error rate on the seven English Open ASR Leaderboard sets against 6.26, better on the 25-language FLEURS average at 10.56 against 11.62, better on long-form TED-LIUM at 2.51 against 2.71, and behind under background noise at 9.04 against 6.72. The licence is CC BY 4.0, the same as the original, so commercial use is allowed with attribution. 9

The deployment number is the reason to care. A 178 MB model that runs at 113x real time on ordinary CPU changes what a transcription feature costs at the margin, and it is the one entry this week that needs no GPU at all.

MiMo-V2.6-Pro-RL is the largest single release of the week and the fastest large-model move in it. Xiaomi's flagship MiMo-V2.6 checkpoint is a sparse mixture of experts with 1.02T total parameters and 42B active, and it takes text, image, video and audio in one model with a 1M-token context. The card reports 71.9 on DeepSWE v1.1 against 74.0 for Claude Opus 5, 73.0 for GPT-5.6 Sol and 70.0 for Claude Fable 5, and describes training as one mixed reinforcement-learning run across coding, general agents, visual work and cybersecurity rather than separate per-domain runs. The licence is MIT, and the weights appeared on September 21. 12

For a builder the property that matters is the context window over repositories, tool traces and multi-session agent runs. The test is one task where that window replaces a retrieval layer you currently maintain, measured on the same inputs.

MiMo-V2.6-Distill-Qwen-9B is the cheap way into the same family: a 9B supervised fine-tune of Qwen/Qwen3.5-9B on MiMo-generated data, covering coding, general agent tasks, visual coding and cybersecurity. Against the Qwen3.5-9B base the card reports SWE Verified at 61.1 against 60.0, SWE Pro at 44.6 against 32.0, AutomationBench at 30.3 against 5.0 and Terminal Bench 2.1 at 37.1 against 27.0. Training used 77.4B tokens, 27.2B of them loss-bearing. The licence is MIT, and the quickstart is an SGLang serve with --reasoning-parser mimo and thinking switched on through the chat template. 30

ThinkingCap-Qwen3.8-27B-GGUF is where the licence changes. The channel covered ThinkingCap on the Qwen3.6-27B base in July, when it shipped under Apache 2.0. This release moves to PolyForm Small Business 1.0.0, which permits commercial use only for a company with fewer than 100 employees and independent contractors and under $1,000,000 of prior-year revenue adjusted for inflation, and adds a personal non-commercial grant from the author. An organisation above that line needs a separate commercial licence. 23

The model itself is a Qwen3.8-27B fine-tune that the card says cuts reasoning tokens by 37% on average while holding 85.8% average accuracy against the base model's 86.6%, and this repository holds the llama.cpp quantizations of it. 22 Reasoning-token cost is the kind of line item worth testing, and the accuracy gap here is small enough to measure against your own eval set. Read the licence first, because the answer decides whether the test is worth running.