HF breakouts Oct 12–19: 127 models clear 10x, while FLUX.3 klein and MOSS-Image passed 600,000 downloads unscored (2026)

Evidence cutoff: October 19, 2026, 06:28 PT. This week's screen asks for two dated readings of a repository's Downloads last month counter, both inside the seven-day window and at least ten times apart.

One hundred and twenty-seven repositories cleared it. A bare ratio test returns 2,200 repositories this week, against 623 a week ago.

That tripling is not a boom. Of the 2,200, 1,558 carry no task tag at all and 1,907 have zero likes, which is the shape of empty or automated repositories rather than shipped models. 1 2

The 5,000-download floor is what makes the number usable. It trims 2,200 to 127, and fourteen of those sit above one hundred thousand downloads. 2

Reading one is a Hub-wide snapshot of model metadata taken on October 13 at 06:29 PT. 1 Reading two is the same file for October 19, taken at 06:28 PT, four minutes before this issue's slot. 2

A third snapshot, from October 12 at 06:40 PT, sits outside the window. Pairing against it instead returns 2,104 on the bare test and 145 above the floor, so the choice of opening reading moves the count by eighteen. 3

Two properties of the counter decide how to read everything below. Downloads last month is a rolling thirty-day sum of file transfers, and the people behind those transfers stay unknown.

Across the 34,568 repositories that had at least a thousand downloads at both readings, the median ratio was 0.989. The snapshots are not drifting against each other. 1 2

Twelve of the week's finalists were re-read against the live Hub counter at publication time, and all twelve matched to the digit. 2

A ratio also cannot score a repository that did not exist when the window opened. This week 115 repositories above 5,000 downloads have no usable earlier reading, 89 of them created inside the window. They get their own table further down, because the week's largest releases are in it.

Repository	Modality	Two dated readings of `Downloads last month`	Change	License and commercial use	Builder fit
`autotrust/JEV-35B`	LLM, decisions	17,426 → 454,857 1 2	26.1x	Apache 2.0, commercial use permitted 4	One pass ranks up to 256 options and returns a probability per option, no prose. 204 ms median on one B200. The NVFP4 sibling needs 23.3 GiB of GPU memory 4
`mradermacher/Simplicio-27B-GGUF`	LLM, coding	977 → 225,544 1 2	230.9x	Apache 2.0, inherited from the base 5	Static GGUF quants of a 27B Qwen-family coding merge, Q4_K_M at 16.9 GB, English and Portuguese. One like on a quarter-million downloads 5
`typhoon-ai/typhoon2.5-qwen3-4b`	LLM, Thai	6,736 → 93,624 1 2	13.9x	Apache 2.0 through the Qwen3-4B base, plus a separate OpenTyphoon terms acceptance 6 7	A Thai instruct model at 4B with 256K context and function calling, built on Qwen3-4B-Instruct-2507 6
`dmis-lab/UltraBERT-380M`	LLM, encoder	101 → 7,520 1 2	74.5x	Apache 2.0 8	A 380M bidirectional encoder with a native 32K context, claiming top marks on both classification and retrieval at its size 8 9
`theforecastingcompany/t0-gamma`	LLM, forecasting	152 → 13,410 1 2	88.2x	Apache 2.0 10	Probabilistic multi-horizon forecasting on covariates, 21 quantile levels, PyTorch and MLX weights. GIFT-Eval CRPS 0.4520, MASE 0.6608 10
`openbmb/MiniCPM-V-4.7-35B-A3B`	Multimodal	630 → 76,410 1 2	121.3x	Apache 2.0 11	The series' first MoE vision model: 35B total, 3B active, claiming 13+ tokens/s at 3 GB of memory on a phone 11
`openbmb/MiniCPM-V-4.7-35B-A3B-gguf`	Multimodal	10,080 → 170,728 1 2	16.9x	Apache 2.0 11	The same model in the format most people actually run, and it outdraws the BF16 repository by 94,318 downloads 2
`openbmb/MiniCPM-V-4.7-1B`	Multimodal	263 → 33,388 1 2	127.0x	Apache 2.0 12	The small sibling, pitched against Qwen3.5 0.8B on both score and inference efficiency 12
`openbmb/MiniCPM-V-4.7-1B-gguf`	Multimodal	2,064 → 53,412 1 2	25.9x	Apache 2.0 12	Also outdraws its own BF16 repository, by 20,024 downloads 2
`vllm-sr/d3-mini`	Multimodal, decisions	1,040 → 33,131 1 2	31.9x	Apache 2.0 13	4.54B including a 0.33B vision encoder; answers choice, yes/no and score questions about text, images or video with a probability each 13
`MCG-NJU/SteadyDancer-14B`	Video, image-to-video	606 → 15,701 1 2	25.9x	Apache 2.0, over the Apache 2.0 Wan2.1-I2V-14B-480P base 14 15	Human image animation that holds the first frame, built to avoid the identity drift of reference-to-video pipelines 14
`DogukanUrker/Qwen-Image-2.1-Turbo-GGUF`	Image generation	5,761 → 162,502 1 2	28.2x	The Qwen research licence, non-commercial, inherited from the base 16 17	Quants of the 8-step distilled 7B DiT, Q6_K at 5.9 GB and an 11.2 GB peak at 1024² on a 12 GB card, with the text encoder kept in system RAM 16
`unsloth/Z-Image-Turbo`	Image generation	5,739 → 68,860 1 2	12.0x	Apache 2.0, unmodified weights from Tongyi-MAI 18 19	The permissive alternative to the Qwen line: same unmodified weights as the upstream 650,482-download release, republished under Apache 2.0 18
`BuzzASR/bulgarian`	Audio, speech recognition	557 → 6,734 1 2	12.1x	MIT 20	A Whisper-large-v3 fine-tune for Bulgarian, reporting CER 1.21 on the combined FLEURS and Common Voice test set against 4.19 zero-shot 20
`jegreene/Swift-1.5-Qwen3.8-27B-heretic-GSQ-RCO-GGUF`	Multimodal, GGUF	10,576 → 237,075 1 2	22.4x	`swift-open-license-1.0`, a named custom licence rather than a standard one 21	An abliterated GGUF of a Qwen3.8-27B merge with MTP heads, self-reporting IFEval 92.98 and GSM8K 97.6 on its own harness 21
`aaronfeller/peptideclm-2-hybrid-base`	LLM, chemistry	1,960 → 30,760 1 2	15.7x	MIT 22	A BERT variant that takes a SMILES string and predicts biochemical properties; one of five repositories from the same unpublished suite 22

LLM

autotrust/JEV-35B is the week's largest verified jump, and it is not a chatbot. It fine-tunes Qwen3.5-35B-A3B into a System 1 decision model that scores up to 256 options in one forward pass. 4

The card reports 59.73 on its own Decision Index 0.3 scoring, a number the authors label as their kit rather than a board entry. 4 It also reports 51.3% on ERQA with no robot data in training, against 52.8 for the best open embodied model in that report. 4

Two things make it shippable. Median latency is 204 ms for a single request on one B200, and the NVFP4 sibling drops the download to 25.6 GB and GPU memory to 23.3 GiB. 4

That places the model on a single 32–48 GB card. For a ranking or triage step that currently calls a chat model, this replaces a per-token bill with one pass.

The card also carries a warning worth reading before you benchmark it: autotrust's model is independent of TypeSafe's hosted Jev, and only the name is shared. 4

mradermacher/Simplicio-27B-GGUF is the counter's puzzle of the week. It gained 230.9x to 225,544 downloads while carrying a single like. 5 The weights are a coding-oriented 27B merge that tags itself qwen and unsloth, quantised from a Portuguese developer's release, Apache 2.0 on both sides. 5 23

A quarter of a million downloads and one like do not describe the same population, and nothing in the data says which is which. Treat the level as discovery, not as an endorsement.

typhoon-ai/typhoon2.5-qwen3-4b is the clearest permissive pick in the text bucket. It is a Thai and English instruct model at 4B with a 256K context and function calling, built on Qwen3-4B-Instruct-2507. 6

The licence field reads Apache 2.0, and the card points that at the Qwen base licence while separately binding users to OpenTyphoon's terms. 6 7 Read the terms before shipping: the Apache 2.0 tag is the model's, not the whole agreement.

dmis-lab/UltraBERT-380M returns to a problem most teams have already solved badly. Sentence encoders tend to win either on classification or on retrieval, not both, and the card claims this one takes both at 380M parameters. 8

Its 32K native context is the part worth testing. The card reports 99% passkey accuracy at every length up to 32K, where 8K encoders degrade. 8 Transformers 5.20 or newer loads it natively; earlier 5.x needs trust_remote_code. 8

theforecastingcompany/t0-gamma is the only forecasting model on the page, and it 88x'd from 152 downloads. 10 It emits 21 quantile levels per horizon, takes known-future covariates, and ships PyTorch and MLX weights. 10

Its published numbers are on public leaderboards rather than its own harness: CRPS 0.4520 and MASE 0.6608 on GIFT-Eval, and a 49.42 skill score on fev-bench. 10 For a demand or inventory forecast, those leaderboards are the right place to start comparing.

Image generation

The image bucket splits cleanly on licence, and that split matters more than any quality number.

DogukanUrker/Qwen-Image-2.1-Turbo-GGUF is a 28.2x mover and the practical way onto Qwen's 8-step distilled 7B DiT. 16 The card documents a 12 GB path: Q6_K at 5.9 GB with the Qwen3-VL text encoder in system RAM peaks at 11.2 GB at 1024². 16

It also inherits the Qwen research licence, which is non-commercial. 17 A quantised copy carries the base model's terms, so a shipping product cannot use it without a separate agreement.

unsloth/Z-Image-Turbo is the other side of that line. It republishes Tongyi-MAI's Z-Image-Turbo weights unmodified under Apache 2.0, and the upstream repository has 650,482 downloads of its own. 18 19

For a product that needs commercial rights and a fast image model, that is the comparison to run, not the Qwen quants.

The week's most useful image artifact has no growth ratio at all. lightx2v/Qwen-Image-2.1-Lightning is a 4-step distilled LoRA for Qwen-Image-2.1, trained with DMD on 16 Huawei Ascend 910B NPUs. 24 It works for both text-to-image and editing, and it appeared on October 13 at 115,621 downloads. 24

Two renders of the same coffee-shop prompt, one from the base model at 40 steps and one from the Lightning LoRA at 4 steps

The same coffee-shop prompt rendered by the Lightning LoRA at 4 steps — The card's own side-by-side: Qwen-Image-2.1 at 40 steps, then the Lightning LoRA at 4 steps, same prompt and seed. The four-step version keeps the chalkboard lettering and the neon sign, and loses some grain in the doorway shadow 24

The quality gap at four steps is small enough that the decision turns on licence again. The Lightning repository declares no licence, and its base is the non-commercial Qwen-Image-2.1. 17 24

Audio

The audio bucket produced two entries, and both are conversions of somebody else's weights rather than new training runs.

Jmica/Drama-3-GGUF is the week's third-largest arrival at 188,725 downloads, and it is an audio.cpp conversion of Fish Audio's Drama 3, a mixture-of-experts speech model with voice cloning and audio editing. 25 The packages run from 22.91 GiB at Q4 experts with an 8-bit backbone down to 15.88 GiB at Q2. 25

The licence is the part to read twice. The GGUF card names the Fish Audio Community License 1.0, while the base repository's own metadata says research licence. 25 26

The linked agreement is version 1.0, dated October 13, and it permits commercial use only while annual revenue stays under USD 500,000. 27

It also withholds hosted service entirely, at any revenue level, unless Fish Audio grants a separate licence. 27 A dubbing studio delivering finished audio is carved out; a self-serve text-to-speech endpoint is not.

BuzzASR/bulgarian is the verified audio entry at 12.1x. It is one of 102 language-specialised Whisper fine-tunes, MIT-licensed, and it reports CER 1.21 on the combined FLEURS and Common Voice set against 4.19 for Whisper-large-v3 zero-shot. 20 28

The pattern is worth noting for anyone building language products: a small fine-tune of a general recogniser still beats the general recogniser on its own language, and the gap here is roughly three and a half times.

Multimodal

OpenBMB's MiniCPM-V 4.7 is the week's centrepiece in any modality. It shipped on October 6 and 7, and thirteen of its repositories cleared the screen this week. 11

The 35B-A3B variant is the first MoE in the MiniCPM-V line: 35B total parameters, 3B active per token, and a claim of 13+ tokens per second at about 3 GB of memory on an iPhone 17 Pro Max. 11

Three techniques carry that claim: lookahead expert scheduling to prefetch the next layer's experts, fine-grained layer-wise quantisation, and DSpark speculative decoding. 11

The architecture detail that reaches your bill is visual token compression. The model compresses tokens inside the vision encoder and lets you pick 4x or 16x compression per request. 11

The card's own evaluation puts 4x compression ahead of 16x almost everywhere, and both ahead of Qwen3.6 35B-A3B on MMBench, at 91.72 and 90.87 against 86.76. 11

Those are the authors' numbers on their own evaluation, in Instruct mode, and the 4x column costs four times the visual tokens. Run your own page types through both settings before picking.

Deployment is unusually broad for a model this size. The card lists vLLM, SGLang, llama.cpp and Ollama for inference, SWIFT and LLaMA-Factory for fine-tuning, and day-0 support for Rockchip, Houmo, Intel, AMD and Arm chips. 11

The 1B variant is the other half of the pitch, aimed at the same efficiency frontier rather than at raw quality. 12 Both sizes ship with a matched DSpark draft head, a 389M-parameter draft that proposes 15 tokens per forward pass. 29

vllm-sr/d3-mini is the second decision model on the page, and it reads images and video as well as text. It is a 4.54B model, including a 0.33B vision encoder, that returns a probability for every option instead of generating an answer. 13

The latency figures are the reason to consider it: 18.1 ms median for a text request, 112 ms with an image, 336 ms with a ten-second video, on one MI325X GPU. 13

It also reports 52.6 on the official Jev Decision Index 0.3.2 text board, first among entries of similar size. 13 Note that this is a board entry, unlike JEV-35B's self-scored figure, so the two numbers are not interchangeable.

MCG-NJU/SteadyDancer-14B is a 2025 research release that resurfaced with a 25.9x week. It animates a person from a still image while holding the first frame, the failure mode that makes reference-to-video pipelines unusable for product shots. 14

It is Apache 2.0 and built on Wan2.1-I2V-14B-480P, which is also Apache 2.0, so the whole stack is commercially usable at 480p. 14 30

Arrivals with no earlier reading

Everything in this section exists, has a licence, and has a download level. None of it has a trajectory, because none of it had a reading when the window opened.

One hundred and fifteen repositories above 5,000 downloads have no usable earlier reading this week. Eighty-nine were created inside the window, and twenty-six were missing from the October 13 file even though they existed. 2

Twenty-two of them sit above 30,000 downloads. Readings, likes and licence tags below all come from the October 19 snapshot. 2

Repository	Oct. 19 reading	Likes	Created	License
`Comfy-Org/MOSS-Image`	673,110	214	Oct. 16	Apache 2.0 31
`black-forest-labs/flux-3-image-klein-comfy`	393,062	874	Oct. 14	FLUX community licence, gated 32
`Jmica/Drama-3-GGUF`	188,725	23	Oct. 13	Fish Audio Community License 1.0 25
`conradlocke/krea2-identity-edit`	153,255	867	Jul. 7	Krea 2 community licence 33
`ml-resources/flux-3-image-klein-comfy`	136,022	5	Oct. 15	FLUX community licence 32
`LiquidAI/LFM2.5-24B-A2B-GGUF`	120,509	212	Oct. 9	LFM 1.0, a named agreement 34
`lightx2v/Qwen-Image-2.1-Lightning`	115,621	90	Oct. 13	none declared; Qwen research base 24
`prism-ml/Ternary-Bonsai-2-9B-gguf`	77,791	101	Sep. 23	Apache 2.0
`nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF`	77,409	97	Oct. 7	OpenMDW-1.1 35
`taronaeo/Granite-4.2-3B-BE-GGUF`	62,768	1	Oct. 18	Apache 2.0
`black-forest-labs/flux-3-image-klein`	62,006	515	Oct. 13	FLUX community licence, gated 32
`black-forest-labs/flux-3-image-klein-base`	56,807	147	Oct. 13	FLUX community licence, gated 32
`lodestones/MOSS-Image-SFT-8-steps-dmd`	45,070	150	Oct. 16	Apache 2.0 36
`unsloth/LFM2.5-24B-A2B-GGUF`	44,125	84	Oct. 13	LFM 1.0 34
`AtomicChat/LFM2.5-24B-A2B-Uncensored-GGUF`	42,767	58	Oct. 9	LFM 1.0 34
`unsloth/LTX-2.5-GGUF`	39,464	10	Oct. 13	LTX-2.x community licence 37
`LiquidAI/LFM2.5-24B-A2B-FP8`	33,158	23	Oct. 12	LFM 1.0 34
`prism-ml/Ternary-Bonsai-2-4B-gguf`	32,571	26	Sep. 23	Apache 2.0
`leemysw/yovoice-model`	32,465	0	Oct. 16	`per-model`, no single document
`unsloth/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-GGUF`	31,719	72	Oct. 14	OpenMDW-1.1 35
`mradermacher/TerminalHorizon-122B-A10B-i1-GGUF`	31,186	1	Oct. 13	Apache 2.0
`dealignai/MiniCPM-V-4.7-35B-A3B-CRACK-GGUF`	30,328	8	Oct. 13	Apache 2.0

Five more sit between 20,000 and 30,000, including a GGUF of the new DeepSeek checkpoint and a Nemotron 3.5 Super VL conversion. 2

Black Forest Labs is the story in that table. FLUX.3 klein appeared on October 13 behind an automatic gate, tagged image-to-image and licensed under the FLUX community licence. 32

Four repositories carry it: the main weights, a base variant, a ComfyUI repackage under the lab's own account, and a third-party repackage. Together they drew 647,897 downloads in six days. 2

The gate is the part to plan around: none of those cards can be read without accepting the terms, so the licence text is a decision you make before you can evaluate the model. 32

MOSS-Image, from Shanghai's OpenMOSS team, is the week's clearest example of a repackaging outdrawing its original. The ComfyUI repackage holds 673,110 downloads; the two upstream repositories hold 2,774 and 3,336. 31 38 39

That is a 110-fold difference on the same weights, and it says where ComfyUI users look for models rather than which model is best. A community 8-step DMD distillation of the same model already has 45,070 downloads of its own. 36

deepseek-ai/DeepSeek-V4.1-Flash-1017 does not reach the table's threshold, and it is the most consequential language model of the week anyway. DeepSeek published this dated checkpoint on October 17 under MIT, and it has 1,558 likes and 17,924 downloads in two days. 40

The card calls it a minor update and then shows large moves: HLE without tools 42.6 against 36.8 for the previous Flash, Terminal-Bench 4.0 at 42.3 against 31.2, DeepSWE v1.1 at 77.0. 40

A MIT-licensed agentic model at that level, two days old, will be worth a fresh reading next Monday. It will have a ratio then.

LiquidAI/LFM2.5-24B-A2B-GGUF at 120,509 and the Nemotron 3.5 Bolt GGUF at 77,409 are a different kind of entry. Both were created before the window opened and are simply absent from the October 13 file, which makes their missing baseline a gap in the data rather than a fact about the models. 2

LFM2.5-24B-A2B is a 24B model with 2B active parameters, built for on-device agents, and it reports 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B. 34 Nemotron 3.5 Bolt packs a 74B vision-language model into 24 GB through native 2-bit quantisation. 35

The rest of the screen

Twenty-three more verified entries sit above 30,000 downloads. Readings come from the two snapshots; licences are the Hub's own tags. 1 2

Repository	Two dated readings	Change	License
`DavidAU/Qwen3.8-27B-Turbo-Brilliance-50X-Reasoning-Instruct-NEO-MAX-MTP-GGUF`	14,787 → 220,057	14.9x	Apache 2.0
`mradermacher/AdaVaR-3B-GGUF`	427 → 186,831	437.5x	Apache 2.0
`mradermacher/Qwen3.8-27B-Fable-5-Coding-Distilled-GGUF`	4,248 → 182,266	42.9x	Apache 2.0
`mradermacher/Qwen3.5-9B-Kimi-k3-Distilled-GGUF`	12,992 → 169,711	13.1x	Apache 2.0
`CompVis/stable-diffusion-v1-1`	5,313 → 168,057	31.6x	CreativeML OpenRAIL-M
`mradermacher/B1-9B-GGUF`	867 → 145,889	168.3x	Apache 2.0
`soyaakinohara/Spark-X2.5-4B-Heretic-jp-gguf`	2,303 → 114,178	49.6x	Apache 2.0
`unsloth/Krea-2-Turbo`	6,275 → 109,143	17.4x	Krea 2 community licence
`sh0wie/Qwen3.8-Flash-Next-REAP-288-GGUF`	2,879 → 106,856	37.1x	Qwen licence
`failspy/Meta-Llama-3-8B-Instruct-abliterated-v3`	9,240 → 99,465	10.8x	Llama 3
`PowCal/Qwen2.5-Math-1.5B-4bit`	866 → 96,349	111.3x	none declared
`rdtand/Qwen3.6-35B-A3B-PrismaQuant-4.75bit-vllm`	5,246 → 90,286	17.2x	Apache 2.0
`lokeshe09/Qwen3.6-35B-A3B-INT8`	48 → 84,108	1,752.3x	Apache 2.0
`RadixArk/Muse-Glimmer-NVFP4`	2,765 → 82,852	30.0x	none declared
`WaveCut/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-NInfer-v3`	855 → 62,074	72.6x	Qwen licence
`timm/densenetblur121d.ra_in1k`	372 → 53,912	144.9x	Apache 2.0
`Qwen/Qwen2.5-Coder-3B-Instruct-AWQ`	3,901 → 51,461	13.2x	Qwen licence
`unsloth/FLUX.2-klein-4B-FP8`	3,686 → 49,736	13.5x	Apache 2.0
`pearsonkyle/Qwen3.8-27B-GPTQ-W4A16`	3,629 → 46,778	12.9x	Apache 2.0
`PowCal/deepseek-math-7b-base-4bit`	108 → 46,326	428.9x	none declared
`unsloth/Z-Image-Turbo-FP8`	2,459 → 37,351	15.2x	Apache 2.0
`WaveCut/Qwen3.8-Flash-Next-ngram-table-NInfer-v3`	546 → 34,042	62.3x	Qwen licence
`sam749/AWPainting-v1-2`	239 → 33,736	141.2x	CreativeML OpenRAIL-M

Ten more sit between 20,000 and 30,000, and 82 between 5,000 and 20,000. 2

Two entries in that table do not belong to this year. CompVis/stable-diffusion-v1-1 was published in August 2022 and gained 31.6x last week, and timm/densenetblur121d.ra_in1k is a 2023 image classifier at 144.9x. 2

Neither had a release. Both sit in evaluation and training pipelines that pull them by name, which is what a download counter sees when a benchmark suite runs.

Before building

Four licence shapes decide most of this week's shortlist, and none of them is visible from the repository's licence tag alone.

The Qwen research licence is the first. Qwen-Image-2.1, its Turbo distill and every quantisation built on them are non-commercial, whatever the derivative's own card says. 17 41

The second is the Fish Audio agreement behind Drama 3, with its USD 500,000 revenue ceiling and its exclusion of hosted speech services at any revenue. 27

The third is the gated FLUX community licence: the terms are accepted before the card can be read, so the licence decision comes before the technical one. 32

The fourth is a set of named agreements that are not standard OSI licences at all: LFM 1.0 for the Liquid models, OpenMDW-1.1 for Nemotron, swift-open-license-1.0 for the Swift derivatives, and per-model for one TTS repository. 21 34 35

Three permissions stand out for a small team. MiniCPM-V 4.7 is Apache 2.0 across all six of its screened repositories, including the GGUF quants. 11 JEV-35B and d3-mini are Apache 2.0 and do one job cheaply. 4 13 Z-Image-Turbo is Apache 2.0 with its weights unmodified. 19

Two counter behaviours will mislead you if you read the numbers without them.

Repackaging beats originals almost everywhere. MOSS-Image shows it most starkly at 673,110 against 6,110 across two upstream repositories, and the Qwen-Image-2.1-Turbo quants show the same pattern, at 162,502 on a third-party GGUF against 25,033 on Qwen's own Turbo repository. 16 31 41

A repackaging also inherits the base model's licence, so the most-downloaded copy of a model is often the one with the tightest terms attached.

Bulk downloading looks like adoption. Five repositories from one account's peptide suite moved together this week, from a combined 2,332 downloads to 136,773, and none of them has a like. 2 The same pattern appears in the GGUF quants that lead the brief table.

None of that proves anything is wrong with the models. It does mean a download count measures transfers, and a like measures people who came back to say something.

Two questions decide which of this week's models survive contact with your product. Does the licence travel with the artifact you would actually download, or only with the repository you found it on. And does the number you are reading come from a repository whose level and whose trajectory you can both explain.

References

1
Hub-wide model snapshot, Oct. 13, 2026
huggingface.co
2
Hub-wide model snapshot, Oct. 19, 2026
huggingface.co
3
Hub-wide model snapshot, Oct. 12, 2026
huggingface.co
4
JEV-35B model card
huggingface.co
5
Simplicio-27B-GGUF model card
huggingface.co
6
typhoon2.5-qwen3-4b model card
huggingface.co
7
OpenTyphoon terms
opentyphoon.ai
8
UltraBERT-380M model card
huggingface.co
9
UltraBERT paper
arxiv.org
10
t0-gamma model card
huggingface.co
11
MiniCPM-V 4.7 35B-A3B model card
huggingface.co
12
MiniCPM-V 4.7 1B model card
huggingface.co
13
d3-mini model card
huggingface.co
14
SteadyDancer-14B model card
huggingface.co
15
Wan2.1-I2V-14B-480P model card
huggingface.co
16
Turbo GGUF model card
huggingface.co
17
Qwen-Image-2.1 model card
huggingface.co
18
Z-Image-Turbo mirror card
huggingface.co
19
Z-Image-Turbo model card
huggingface.co
20
BuzzASR Bulgarian model card
huggingface.co
21
Swift-1.5 GGUF model card
huggingface.co
22
peptideclm-2 hybrid model card
huggingface.co
23
Simplicio-27B model card
huggingface.co
24
Lightning LoRA model card
huggingface.co
25
Drama-3 GGUF model card
huggingface.co
26
Drama 3 model card
huggingface.co
27
Fish Audio Community License 1.0
huggingface.co
28
BuzzASR paper
arxiv.org
29
MiniCPM-V 4.7 35B-A3B DSpark model card
huggingface.co
30
Wan2.1-I2V-14B-480P model card
huggingface.co
31
MOSS-Image ComfyUI repackage
huggingface.co
32
flux-3-image-klein repository
huggingface.co
33
krea2-identity-edit repository
huggingface.co
34
LFM2.5-24B-A2B model card
huggingface.co
35
Nemotron 3.5 Bolt model card
huggingface.co
36
MOSS-Image 8-step DMD card
huggingface.co
37
LTX-2.5 GGUF model card
huggingface.co
38
MOSS-Image-SFT model card
huggingface.co
39
MOSS-Image-Base model card
huggingface.co
40
DeepSeek-V4.1-Flash-1017 model card
huggingface.co
41
Qwen-Image-2.1-Turbo model card
huggingface.co