HF breakouts Sep 28–Oct 5: 75 models clear 10x, led by an Ornith draft checkpoint at 10,903x
Seventy-five Hugging Face repositories cleared the >10x download screen this week, led by Ornith's DFlash draft checkpoints and two AutoTrust decision models, with the licence terms and builder fit of fourteen of them spelled out.
Evidence cutoff: October 5, 2026, 09:00 PT. Seventy-five Hugging Face repositories cleared this week's screen, which asks for two dated readings of the same repository's
Downloads last month counter, both taken inside the seven-day window, at least ten times apart. The widest multiple sits on a draft checkpoint whose only job is to make a larger model decode faster. The largest absolute move is 778,377 downloads, on a two-class image classifier.That count comes from a wider sweep than archived model pages alone can support. A bare ratio test across the Hub — any repository standing at 1 or more on September 29 that ended the window at least ten times higher — returns 768 repositories, and a traction floor of 5,000 downloads at the cutoff leaves 75. 1
Fourteen of the 75 take a full row below, with the licence read to the clause and the builder angle. Nineteen more that reached 20,000 downloads are listed in brief after the modality sections. The remaining 42 sit between 5,000 and 20,000, in a band where quantizations, abliterations, merges and MLX or ComfyUI repackagings of the models named here sit beside older and narrower ones.
Reading 1 for every row is a Hub-wide snapshot of model metadata taken at 6:04 a.m. PT on September 29 and published on the Hub; it records each repository's rolling 30-day figure at that moment. Reading 2 is the same figure, read from the repository page at the cutoff. Both appear in every row, so the multiple can be recomputed rather than taken on faith. Where the Internet Archive holds a capture of the model page from inside the window, that reading is cited beside the row as an independent check; six rows carry one, and each agrees with the snapshot once the hours between the two observations are allowed for. 1
Two properties of the counter shape how these numbers read. Nine of the fourteen repositories are younger than 30 days, and for those the Hub's all-time counter equals the monthly one, so the figure is the whole life of the repository rather than a rolling month. 1 On five rows the first reading is a base in single digits — the two Ornith draft checkpoints, JEV-9B, JEV-27B and OpenDecider-nano — so the multiple is large because the starting number is small.
| Repository | Modality | Two dated readings of Downloads last month | Change | License and commercial use | Builder fit |
|---|---|---|---|---|---|
autotrust/JEV-9B | LLM, typed decisions | 471 (Sept. 29, 6:04 a.m. PT) → 305,502 at cutoff; 336 on an archived page from Sept. 28 2 | 648.6x | Apache 2.0. Commercial use permitted 3 | A dual-head fine-tune of Qwen3.5-9B: one head returns a calibrated choice or score for a typed question in a single forward pass, the other generates normally. A vision path for screenshots and camera frames landed on October 3 3 |
autotrust/JEV-27B | LLM, typed decisions | 205 (Sept. 29, 6:04 a.m. PT) → 134,153 at cutoff; 120 on Sept. 28 and 785 on Oct. 2, both on archived pages 45 | 654.4x | Apache 2.0 on the fine-tune, over the Apache 2.0 Qwen3.8-27B base 6 | The larger sibling on Qwen3.8-27B, distilled from the closed TypeSafe Jev 1.13 teacher; the card reports KL 0.0186 and a choice top-1 agreement of 0.903 across test rows, 0.958 where the teacher was decisive 6 |
manjunathshiva/opendecider-nano | LLM, typed decisions | 71 (Sept. 29, 6:04 a.m. PT) → 6,057 at cutoff; the same 71 on an archived page from Sept. 29 and 513 from Oct. 3 78 | 85.3x | Apache 2.0 9 | A ~400M encoder built on Ettin-400M that returns a calibrated probability per option for choice, score and noul questions; the card measures 17 ms per call on an L40S and 18 ms on an M4 Max 9 |
ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF | LLM, compression | 19,505 (Sept. 29, 6:04 a.m. PT) → 389,398 at cutoff; 351,230 on an archived page from Oct. 4 10 | 20.0x | The card declares Apache 2.0 on the compressed build; the Qwen3.8-Flash-Next base carries the Qwen Community License 1.0, which allows commercial use and requires a separate Qwen licence for a model-as-a-service or AI-coding-assistant business 11 | Half of the base model's 512 experts per layer removed outright and the rest stored at 3.5 bits per weight, which brings a 176.9B-parameter model to a 29.6 GB resident working set. The card measures 75.60 on SWE-bench Verified against 82.80 for the unpruned base 12 |
autotrust/GLM-5.3-Flash-GGUF-DGX-Spark | LLM, compression | 439 (Sept. 29, 6:04 a.m. PT) → 31,477 at cutoff; 28,936 on an archived page from Oct. 4 13 | 71.7x | MIT on the compressed build and MIT on the zai-org/GLM-5.3-Flash base. Commercial use permitted 14 | 32 of 288 routed experts per layer removed and the survivors quantized, taking the GGUF from 90 GiB to 79.1 GiB so it fits one 128 GB DGX Spark with enough memory left for 64K context 14 |
prithivMLmods/Gender-Classifier-Mini | Image, classification | 36,777 (Sept. 29, 6:04 a.m. PT) → 778,377 at cutoff | 21.2x | Apache 2.0. Commercial use permitted 15 | A SigLIP2-base fine-tune for one binary label, scoring 0.972 accuracy and F1 on its own 5,000-image test set. The card states the task, the accuracy and the licence, and stops there 15 |
Noctaluna/Noct-Q-Uncensored-Qwen-Image-2.1 | Image generation | 401 (Sept. 29, 6:04 a.m. PT) → 37,070 at cutoff | 92.4x | Qwen Research License, the same agreement the base model ships under: non-commercial purposes only 16 | An int8 merge of Qwen-Image-2.1 for ComfyUI, shipped with a workflow and an 8–12 GB card target, aimed at explicit imagery the base model refuses. The licence travels with it whatever the merge changes 17 |
QuantFunc/Minimax-H3-Quantfunc-4bit | Video generation | 1,671 (Sept. 29, 6:04 a.m. PT) → 28,828 at cutoff | 17.3x | MiniMax H3 Community License: the grant covers worldwide use outside the European Union, the United Kingdom, South Korea and the United States, needs MiniMax's separate licence above US$20M in yearly revenue, and requires the name "MiniMax H3" on the interface of any commercial product that uses it 18 | A 4-bit SVDQuant of MiniMax H3's core weights, cutting them to a quarter of their size for ComfyUI with the native video and audio generation path intact 19 |
vasista22/whisper-tamil-large-v2 | Audio, speech recognition | 402 (Sept. 29, 6:04 a.m. PT) → 35,015 at cutoff | 87.1x | Apache 2.0. Commercial use permitted 20 | A Whisper large-v2 fine-tune for Tamil, reporting 7.5 word error rate on FLEURS ta_in and 6.61 on Common Voice 11 — the kind of number a monolingual product needs before it ships its own transcription 20 |
vadimbelsky/qwen3-asr-arabic-uae | Audio, speech recognition | 675 (Sept. 29, 6:04 a.m. PT) → 23,661 at cutoff | 35.1x | Apache 2.0 on the fine-tune; the base is the Apache 2.0 Qwen3-ASR-1.7B 21 | Qwen3-ASR-1.7B fine-tuned on roughly 22,500 Emirati Arabic samples, which the card measures at 9.98% word error rate against 13.53% for the base on the same 2,497-item validation set 21 |
mlx-community/chatterbox-turbo-4bit | Audio, text to speech | 419 (Sept. 29, 6:04 a.m. PT) → 5,147 at cutoff | 12.3x | Apache 2.0 as declared on this conversion; the ResembleAI/chatterbox-turbo base is MIT. Both permit commercial use 22 | A 4-bit MLX conversion of Chatterbox Turbo with voice cloning and inline emotion tags such as [sigh] and [chuckle], which puts expressive synthesis on Apple Silicon 22 |
ornith-ai/Ornith-1.5-35B-A3B-DFlash | Multimodal, inference acceleration | 9 (Sept. 29, 6:04 a.m. PT) → 98,127 at cutoff; the counter was still unpopulated on Sept. 28 and read 41,854 on Oct. 2 2324 | 10,903x | MIT on this checkpoint and MIT on the Ornith-1.5-35B-A3B base, which was the licence question hanging over this family in August 25 | A block-diffusion draft model for speculative decoding: it proposes blocks of tokens in parallel and the target Ornith model verifies them. It runs only alongside the target model, which serves the actual requests 25 |
ornith-ai/Ornith-1.5-9B-DFlash | Multimodal, inference acceleration | 8 (Sept. 29, 6:04 a.m. PT) → 64,730 at cutoff | 8,091x | MIT on this checkpoint and on the Ornith-1.5-9B base 26 | The same draft-and-verify arrangement for the 9B member of the family, published with its serving configuration for the DFlash-enabled inference stack 26 |
ibm-granite/granite-4.0-3b-vision | Multimodal, document extraction | 678 (Sept. 29, 6:04 a.m. PT) → 37,208 at cutoff | 54.9x | Apache 2.0. Commercial use permitted 27 | A 3.5B language model with 0.5B of LoRA adapters for document work: charts to CSV, summary or code, complex tables to JSON, HTML or OTSL, and key-value extraction across varied layouts. The base model keeps serving text-only requests with the adapter unloaded 27 |
LLM
Three of the five rows in this section are the same contraption: a small model that answers a typed question in one forward pass and returns a label instead of prose. This is the third week in a row one of them has cleared the screen, after Mapika/decider-2b on September 21 and kev-9b alongside GLiNER2.5-Decide a week later. What is new is the teacher:
autotrust/JEV-9B and its successor are students of a closed model, and both cards report how closely the student matches it.JEV-9B carries two heads on one set of Qwen3.5-9B weights. One returns a calibrated probability for every option in a question of the
choice, score or noul type; the other is the untouched language model. A request routes to whichever the caller wants, so a classifier and a generator share one vLLM engine. The card reports a noul AUROC of 0.996, a noul Brier score of 0.0013, and a top-1 agreement with the teacher of 0.903 across all test rows. 3 On October 3 the repository added vision, and the card documents what that buys: a simulated robot arm that answers two questions per camera frame and completes a pick-and-place on 10 of 20 random scenes at about 165 ms per decision, and a headless-browser agent that completes 95% of 60 multi-step tasks at about 0.2 s per click. 3The repository is twelve days old, so its all-time download count and its monthly one are the same number. 1
manjunathshiva/opendecider-nano reaches the same kind of answer from a 400M encoder rather than a language model, at a reported 17 ms per call on an L40S and 9 ms per question when batched. Its card puts it ahead of Laya's typed-decision checkpoint on the same benchmark, 0.796 against 0.766, with both fine-tuned on identical train splits. 9Those accuracy and latency figures are the authors' own, measured on the authors' suites. The number that decides anything for you is agreement with the routing or triage decision your product already runs, counted over your own logged traffic.
The other two rows are compressions of large mixtures of experts, and they arrive at the same trick from opposite ends. ISTA-DASLab removed half of
Qwen3.8-Flash-Next's 512 experts per layer and stored what remained at 3.5 bits per weight, which brings a 176.9B-parameter model to a 29.6 GB resident working set on one 32 GB accelerator. Code generation survives it better than agentic work: 98.7% of the base score on LiveCodeBench v6 against 91.3% on SWE-bench Verified. 12Read two licences on that one. The card declares Apache 2.0 for the compressed build; the base model carries the Qwen Community License 1.0, which grants commercial use and then requires a separate Qwen licence if you run a model-as-a-service or an AI coding assistant business. Internal use is carved out. 11
AutoTrust's second compression entry is the lightest of the week: 32 of 288 routed experts removed per layer, the survivors quantized hard, and the GGUF drops from 90 GiB to 79.1 GiB. That 11 GiB is the difference between the 744B build's memory profile and this one on a single 128 GB DGX Spark, and it is what pays for 64K of context on the machine. MIT covers both the build and the base. 14
Image generation
The week's largest absolute move belongs to the smallest thing on the list.
prithivMLmods/Gender-Classifier-Mini is a SigLIP2-base fine-tune that sorts an image into one of two classes, and its counter went from 36,777 to 778,377. The card reports 0.972 accuracy and F1 on its own 5,000-image test set and states its licence; its only reported test is that same set, and the counter records downloads rather than the people behind them. A two-class demographic label is the kind of output that invites a check against your own jurisdiction and your own users before it goes into a product. 15The image bucket has been Qwen-Image-2.1 derivatives for three weeks running, and this week's entry changes the output rather than the packaging.
Noctaluna/Noct-Q-Uncensored-Qwen-Image-2.1 is an int8 merge of the base model, shipped with a ComfyUI workflow and profiles for 8–12 GB cards, tuned for the explicit material the base model declines. The licence is the part that matters: the repository publishes the Qwen Research License Agreement itself, dated September 20, 2026, and that agreement limits use to non-commercial purposes. The same terms govern the base model and every merge, quant and LoRA built on it. 16QuantFunc/Minimax-H3-Quantfunc-4bit is the one video entry, a 4-bit SVDQuant of MiniMax H3's core weights that cuts them to a quarter of their size for ComfyUI while keeping the native video-and-audio path intact. Its licence is the most conditional of the fourteen rows. The MiniMax H3 Community License covers use worldwide outside the European Union, the United Kingdom, South Korea and the United States, so a builder shipping into any of those four territories sits outside the grant. Above US$20M in yearly revenue it needs a separate agreement from MiniMax, and any commercial product using the weights has to display "MiniMax H3" on its interface. 1819Audio
Two of the three audio rows are single-language speech recognition fine-tunes with a measured word error rate, which is the most directly testable claim on this page.
vasista22/whisper-tamil-large-v2 is a Whisper large-v2 trained on Tamil corpora, reporting 7.5 word error rate on FLEURS ta_in and 6.61 on Common Voice 11. 20 vadimbelsky/qwen3-asr-arabic-uae is Qwen3-ASR-1.7B fine-tuned on about 22,500 Emirati Arabic samples, which the card measures at 9.98% word error rate against 13.53% for the base model on the same 2,497-item validation set. 21 Both are Apache 2.0, and both exist because a general-purpose model handles a specific dialect or script worse than the people using it can accept.The third is a packaging win.
mlx-community/chatterbox-turbo-4bit brings ResembleAI's Chatterbox Turbo to Apple Silicon in 4-bit MLX form, with voice cloning from a reference clip and emotion tags written straight into the input text. For a Mac-based product the practical question is whether the reference clip you need per voice makes the clone cheap enough, and that is a test rather than a download count. 22Multimodal
The widest multiple of the week comes from a repository that holds no model anyone can talk to.
ornith-ai/Ornith-1.5-35B-A3B-DFlash is a block-diffusion draft checkpoint: it proposes several tokens at once, and the Ornith model it is paired with accepts or rejects them. The card says plainly that this is a companion checkpoint for a DFlash-enabled serving setup, paired with the Ornith model it accelerates. Its value appears only under a DFlash-enabled serving setup. 25The family itself is a follow-up. This channel put Ornith 1.5 on the watchlist on August 24 at a 1.38x increase, with the licence unstated on the card. Three DFlash checkpoints for the 1.5 family appeared on September 20 and all three cleared this screen — the third, for the 397B member, moved from 7 to 11,379 — and every one of them declares MIT, as do the Ornith-1.5-35B-A3B and Ornith-1.5-9B bases they attach to. 26
What a draft checkpoint tells a builder is narrower than its download count suggests. Its value is a throughput number on the exact target model and the exact serving stack you run, so the measurement is tokens per second for the pair against tokens per second for the target alone, on your prompts. The counters here, 9 and 8 at the start of the window, mean the multiple is arithmetic on a tiny base.
ibm-granite/granite-4.0-3b-vision is the opposite kind of entry: a small model with a narrow job and a long history. It is a 3.5B Granite 4.0 Micro base with 0.5B of LoRA adapters trained for document extraction — charts to CSV, summary or code, complex tables to JSON, HTML or OTSL, and key-value pairs pulled from varied layouts. The base keeps answering text-only requests with the adapter unloaded, so one deployment covers both. The card points to the ChartNet paper behind the training data and to a newer 4.1 release in the same line. 27The rest of the screen
Nineteen more repositories reached 20,000 downloads at the cutoff and cleared the ratio. Two of them are the largest absolute movers outside the table:
YalaLab/Pillar0-ChestCT, gated behind a name-and-organisation request, with an ECL-2.0 licence and an unpublished card, and HooshvareLab/bert-base-parsbert-uncased, a Persian BERT that has been on the Hub since 2022, with an empty licence field.Each reading in that table comes from the September 29 snapshot and the live counter at the cutoff, the same two sources as the rows above. 1 The band below it, the 42 repositories between 5,000 and 20,000, holds a similar spread: quantizations and abliterations of models already named here, a handful of narrow single-purpose models, and older checkpoints that moved for reasons a download figure leaves open.
Before building
- Run each candidate on a task from your own product, with the pass-fail threshold fixed before the first run rather than after it.
- Measure peak memory, throughput and hosting cost on the machine you intend to ship on. The GSQ-RCO working set and the DGX Spark footprint are the authors' measurements on their hardware.
- Read the repository licence and the base model's licence, and follow the stricter one. Three of this week's fourteen carry a condition that changes the answer: Noct-Q is non-commercial through its Qwen-Image-2.1 parent, MiniMax H3 excludes four territories and needs a separate agreement above US$20M a year, and the pruned coder inherits the Qwen Community License requirement of a separate Qwen licence for a hosting or AI-coding-assistant business.
- Treat one counter reading as a level. Direction appears only when two dated readings of the same repository are compared, and this issue prints both for every row so you can recompute them.
References
- 1Hub-wide model snapshot, Sept. 29, 2026
huggingface.co
- 2Archived JEV-9B model page, Sept. 28, 2026
web.archive.org
- 3JEV-9B model card
huggingface.co
- 4Archived JEV-27B model page, Sept. 28, 2026
web.archive.org
- 5Archived JEV-27B model page, Oct. 2, 2026
web.archive.org
- 6JEV-27B model card
huggingface.co
- 7Archived OpenDecider-nano page, Sept. 29, 2026
web.archive.org
- 8Archived OpenDecider-nano page, Oct. 3, 2026
web.archive.org
- 9OpenDecider-nano model card
huggingface.co
- 10Archived GSQ-RCO Coder page, Oct. 4, 2026
web.archive.org
- 11Qwen Community License 1.0, Qwen3.8-Flash-Next
huggingface.co
- 12GSQ-RCO Coder model card
huggingface.co
- 13Archived GLM-5.3-Flash DGX Spark page, Oct. 4, 2026
web.archive.org
- 14GLM-5.3-Flash-GGUF-DGX-Spark model card
huggingface.co
- 15Gender-Classifier-Mini model card
huggingface.co
- 16Qwen Research License Agreement, Noct-Q repository
huggingface.co
- 17Noct-Q model card
huggingface.co
- 18MiniMax H3 Community License Agreement
huggingface.co
- 19Minimax-H3-Quantfunc-4bit model card
huggingface.co
- 20Whisper Tamil Large-v2 model card
huggingface.co
- 21Qwen3-ASR Arabic — UAE model card
huggingface.co
- 22chatterbox-turbo-4bit MLX model card
huggingface.co
- 23Archived Ornith-1.5-35B-A3B-DFlash page, Sept. 29, 2026
web.archive.org
- 24Archived Ornith-1.5-35B-A3B-DFlash page, Oct. 3, 2026
web.archive.org
- 25Ornith-1.5-35B-A3B-DFlash model card
huggingface.co
- 26Ornith-1.5-9B-DFlash model card
huggingface.co
- 27Granite 4.0 3B Vision model card
huggingface.co
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.