Neodrop
Create
FeedDiscoverMy ChannelsAuto-publishJoin Our Discord
FeedDiscoverCreateMy Channels
HF breakouts Oct 5–12: 115 models clear 10x, and a 27B in under 8 GB drew 242,329 downloads in five days
2026-10-12

HF breakouts Oct 5–12: 115 models clear 10x, and a 27B in under 8 GB drew 242,329 downloads in five days

One hundred and fifteen Hugging Face repositories cleared the >10x download screen for October 5–12, with the licence terms and builder fit of fifteen of them spelled out, plus thirteen repositories that arrived too recently to pair.

Hugging Face Surging Models
Hugging Face Surging Models@NeoDrop Official
Evidence cutoff: October 12, 2026, 09:00 PT. This week's screen asks for two dated readings of a repository's Downloads last month counter, both taken inside the seven-day window and at least ten times apart.
One hundred and fifteen repositories cleared it. A bare ratio test returns 623 repositories on the October 6 pair and 922 on the October 5 pair.
The traction floor does the work. Five thousand downloads at the cutoff trims the 623 to 115, and thirty-three of those sit above twenty thousand. 12
Reading one is a Hub-wide snapshot of model metadata published on the Hub, taken on October 6 at 6:38 a.m. PT. It records every repository's rolling 30-day figure at that moment. 1
Reading two is the same snapshot for October 12, taken at 5:40 a.m. PT. 2 Its figure matched the live counter on every one of the thirty-one repositories checked against it.
Two properties of the counter shape every number here. Downloads last month is a rolling 30-day sum, so it counts file transfers over a month, and the people behind those transfers stay unknown.
For a repository younger than thirty days, the Hub's all-time figure equals the monthly one. The monthly figure is then the whole life of the repository.
Each snapshot holds about 3.1 million repositories, and their contents differ slightly. Eighteen repositories that existed before the window opened and now sit above 5,000 downloads are missing from the October 6 reading.
Most of the widest multiples this week sit on builds of somebody else's weights: quantisations, GGUF packages, MLX conversions and merges. Fifteen of the 115 take a full row below. Twenty-four more above twenty thousand downloads are listed in brief after the modality sections.
RepositoryModalityTwo dated readings of Downloads last monthChangeLicense and commercial useBuilder fit
Alennnndy/Qwen3.6-27B-AWQ-4bitLLM, compression516 → 104,034 12201.6xApache 2.0 on the build and on the Qwen3.6-27B base 34AWQ W4A16 weights take the model from about 54 GB to about 18 GB, with the vision layers held at full precision and a vLLM path 3
melcheikh/gemma-4-31B-it-qat-NVFP4-BlackwellLLM, compression362 → 94,994 12262.4xThe card declares Apache 2.0, the Hub's licence tag reads gemma, and the card's own licence link points at Google's Gemma 4 terms. Commercial use needs the Gemma 4 document read in full 56NVFP4 from Google's unquantised QAT weights, for native FP4 on Blackwell under vLLM 0.20 or newer. A newer MSE-scaled variant sits in a sibling repo 5
empero-ai/Qwen3.8-4B-DistillLLM, distillation8,941 → 114,190 1212.8xApache 2.0, inherited from the Qwen3.5-4B base. Commercial use permitted 7A reasoning distillation onto Qwen3.5-4B. The card reports MMLU at 0.553 against 0.354 for the base and GSM8K at 0.785 against 0.850, so mathematics slips while breadth gains 7
bleysg/Qwen3.5-122B-A10B-int4-fp8-hybridLLM, compression6,104 → 95,086 1215.6xThe licence field reads other and names the Qwen licence; the build inherits the Qwen3.5-122B-A10B and Intel AutoRound terms of its two parents 89Routed experts and attention stay INT4 while the shared experts go to FP8, for a 128 GB DGX Spark. It serves under one project's patched vLLM; stock vLLM refuses the mixed scheme 8
JackKozmo29/sevclassifier-jev-style-1.5bLLM, typed decisions430 → 29,381 1268.3xApache 2.0 on all four guard models and on the Qwen2.5-1.5B-Instruct base 10A 1.5B fine-tune that returns one JSON object per incident: severity, escalation flag, blast radius. The card reports 92.9% F1 against 93.6% for the hosted service it copies 10
onebrain-ai/onebrain-rerank-v1LLM, reranking174 → 11,097 1263.8xApache 2.0, following the bge-reranker-v2-m3 base 1112A 569 MB int8 ONNX build of bge-reranker-v2-m3, shipped inside a CLI and pinned by checksum. The card states that v1 ships as a size and speed optimisation, with fine-tuning left to later versions 11
CypherTroopers/CPH-Community-7BLLM, domain131 → 78,603 12600.0xMIT 13A QLoRA fine-tune of Qwen2-7B for Cypherium node operation, validator troubleshooting and RPC work, shipped as Q4_K_M for CPU-only llama.cpp on a 2–6 vCPU box 13
mohsenvand/mindview-t2i-turboImage generation223 → 9,716 1243.6xApache 2.0 across every part it names: Ternary Bonsai 1.7B, Bonsai Image 4B over FLUX.2 klein 4B, and the three trained branches. TAEF2 is MIT 14One 1,208 MB file painting 512×512 in one, two or four steps inside a browser on WebGPU, with the computation running locally. Of its about 4.61 billion weights, 165 million sit off the ternary grid 14
mohammedaly22/VoiceTut-TTS-GGUFAudio, text to speech628 → 7,661 1212.2xApache 2.0 15Standalone GGUF packages of an OmniVoice-based model for Egyptian Arabic and Arabic–English code-switching, run without Python. The card measures up to 6.6× the PyTorch package on the same GPU 15
mohammedaly22/QwenCleo-ASR-GGUFAudio, speech recognition609 → 7,568 1212.4xApache 2.0 16The transcription half of the same pair: a Qwen3-ASR-1.7B fine-tune for Egyptian Arabic and Arabic↔English code-switching, packaged for the same non-Python runtime 16
team-lucid/hubert-large-koreanAudio, speech recognition245 → 6,014 1224.5xApache 2.0 17A Korean speech recogniser built on Facebook's HuBERT large, trained on Cloud TPUs through Google's research programme. The card documents the architecture, and the number that matters for your product is the one you measure on your own audio 17
Alibaba-DAMO-Academy/RynnBrain1.1-2BMultimodal, embodied4,927 → 64,279 1213.0xApache 2.0 on the 2B checkpoint and the Qwen3.5-2B base 18Alibaba DAMO's embodied vision-language model, released across 2B, 9B and 122B-A10B on July 16. It adds 3D grounding and a contact-point task to image-plane localisation 1819
mindchain/imajev-9b-GGUFMultimodal, typed decisions772 → 10,326 1213.4xApache 2.0 20A vision decision model on Qwen3.5-9B, shipped as eleven GGUF quantisations from Q3_K_S to Q8_0. Q5_K_M lands at 7.44 GiB, which fits an 8 GB card 20
mlx-community/olmOCR-2-7B-1025-mlx-8bitMultimodal, document parsing370 → 34,595 1293.5xApache 2.0, over the Apache 2.0 Qwen2.5-VL-7B-Instruct base 2122Allen AI's olmOCR-2 in 8-bit MLX form for Apple Silicon, loaded through mlx-vlm. Read the original card for what the model does; the conversion adds the MLX packaging 21
hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2vVideo generation1,609 → 20,813 1212.9xother, with an empty licence link on the card 23A Diffusers pipeline for HunyuanVideo 1.5 at 720p text-to-video, with variable-length attention masks and prebuilt kernel backends. 121 frames at 50 steps per clip 23

LLM

Two of the seven text rows compress a large mixture-of-experts model, and each raises a different obstacle between the download and the running server.
Alennnndy/Qwen3.6-27B-AWQ-4bit holds Qwen3.6-27B's weights at four bits with 16-bit activations, calibrated on 512 samples from HuggingFaceH4/ultrachat_200k. 3 The card estimates about 18 GB of VRAM against about 54 GB in FP16. 3 The vision encoder stays at full precision, so the multimodal path survives. Apache 2.0 covers both the build and the base. 4
bleysg/Qwen3.5-122B-A10B-int4-fp8-hybrid splits the quantisation by role. Routed experts and attention stay INT4; the dense shared experts, which every token reads, become calibrated FP8. 8 The card measures the change at 28.2 → 36.0 tokens per second on a DGX Spark. 8 Stock vLLM refuses that mixed scheme, so the build needs one project's patched engine. 8
melcheikh/gemma-4-31B-it-qat-NVFP4-Blackwell repacks Google's unquantised QAT weights into NVFP4 for Blackwell cards. 5 Read the licence before the model. The card declares Apache 2.0; the Hub's licence tag on the same repository reads gemma; and the card's own licence link opens Google's Gemma 4 terms. 56 Treat the Gemma 4 document as the governing text and the Apache 2.0 field as the weaker of the two.
empero-ai/Qwen3.8-4B-Distill is the week's largest absolute gain inside the table, and its card is unusually frank about the trade. MMLU rises to 0.553 from 0.354 on the base model, and GSM8K falls to 0.785 from 0.850. 7 The build also needs the Gated DeltaNet kernels installed, or the linear-attention layers fall back to slow PyTorch paths. 7
A fourth week of typed-decision models arrives as a family. Four repositories under one account fine-tune Qwen2.5-1.5B-Instruct to return a single JSON object instead of prose: sevclassifier for incident severity, depguard, mailguard and codeguard for software supply chain, email and code review decisions. 10242526 Together they went from 3,395 downloads on October 6 to 120,047 at the cutoff. 12 Every card compares itself with TypeSafe's hosted Jev service, and each reports a small accuracy gap in exchange for keeping client data on your own machine. 10 All four cards show enterprise-ai-lab/ as the repository path in their sample code, while the account that published them is JackKozmo29.
onebrain-ai/onebrain-rerank-v1 takes the opposite route through the same idea. It carries the bge-reranker-v2-m3 weights exported to ONNX at dynamic int8, 569 MB in one file, with no fine-tuning of its own. 11 For a first-stage retriever that already returns ten candidates, that is a drop-in second stage with no Python dependency and a checksum to pin.
CypherTroopers/CPH-Community-7B is the narrowest model on the page: Qwen2-7B fine-tuned for one blockchain's node operation, shipped as Q4_K_M for CPU-only llama.cpp. 13 Six hundred times the October 6 reading on a base of 131 shows what a small, specific audience looks like in a counter.

Image generation

The image bucket produced one qualifier this week, and it is the most unusual file on the page. mohsenvand/mindview-t2i-turbo is one 1,208 MB file that paints a 512×512 picture in a browser on WebGPU, with the computation running locally. 14 Of its 4.61 billion weights, all but 165 million sit at −1, 0 or +1. 14
Eight rows of generated images in three columns labelled one step, two steps and four steps
Eight prompts at one, two and four steps, published on the model card. The lettering softens in the first column, and the seventh row paints its sign twice in the second 14
The card reports about four seconds per picture at one step on an M4 MacBook Air, six at two and fifteen at four, and adds that the fanless machine slows by about a third after minutes of work. 14 Every component it names is Apache 2.0, including the ternary rework of FLUX.2 klein 4B at the centre of it. 14
The rest of the image bucket stayed on the shelf. Qwen-Image-2.1 derivatives have cleared this screen in each of the last three issues, and this week's Qwen-Image-2.1-Turbo builds are listed among the repositories that arrived after the window opened.

Audio

Two of the three audio rows are one person's work on one dialect, packaged for a runtime that has no Python in it. mohammedaly22/VoiceTut-TTS-GGUF and its companion mohammedaly22/QwenCleo-ASR-GGUF cover Egyptian Arabic and Arabic↔English code-switching, the mixture a speaker actually produces mid-sentence. 1516 Both cards measure three to six times the speed of the PyTorch package on the same GPU. 15
A general-purpose recogniser mishandles a dialect often enough that the people who speak it build their own. That is the whole case for team-lucid/hubert-large-korean, a 2023 Korean model on Facebook's HuBERT large that resurfaced this week. 17 Its card documents the encoder and publishes no word error rate, so the first useful number is the one from your own audio. 17

Multimodal

Alibaba-DAMO-Academy/RynnBrain1.1-2B is the largest research-lab entry this week. Alibaba's DAMO Academy released the RynnBrain 1.1 family on July 16 at 2B, 9B and 122B-A10B, all trained on one recipe for embodied work. 18 Version 1.1 adds metric 3D grounding and a contact-point prediction task, and the companion work translates that understanding into robot control. 19
The typed-decision line from the LLM section continues into vision. mindchain/imajev-9b-GGUF puts the decision model on Qwen3.5-9B, adds an image path, and ships eleven GGUF quantisations. 20 The card's own demonstration is the one worth repeating: the same page, then the same page blackened, moves the answer distribution from 0.482 on one option to 0.887 on another, which shows the image reached the forward pass. 20 Its Q5_K_M build fits an 8 GB card at 7.44 GiB. 20
Document parsing returned in MLX form. mlx-community/olmOCR-2-7B-1025-mlx-8bit converts Allen AI's olmOCR-2 to 8-bit MLX for Apple Silicon, with Apache 2.0 inherited from the Qwen2.5-VL-7B-Instruct base. 21 Its own card forwards the original and adds the packaging, so evaluate it on your own worst documents. 22
hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v is the one video row, a Diffusers pipeline for 720p text-to-video at 121 frames and 50 steps per clip. 23 Its licence field reads other with an empty link, so the weights' terms stay unresolved on the Hub. 23

Repositories that did not exist when the window opened

Thirteen repositories above twenty thousand downloads were created after the window opened, which leaves them without an earlier reading to divide by. 2 Seventy-seven more crossed five thousand the same way. Their readings are levels, and the licence column is the part worth reading.
RepositoryOct. 12 readingLikesCreatedLicense
ConwayResearch/Underdog-Saluki-27B-1.0242,329348Oct. 7Apache 2.0 27
AtomicChat/Qwen-Image-2.1-Turbo-Uncensored-GGUF161,715359Oct. 9Apache 2.0 declared 28
kindlingai/glm-5.3-exlr8-k2-k3-k4117,64912Oct. 8Licence reads glm-5.3 29
unsloth/Qwen-Image-2.1-Turbo-GGUF108,48536Oct. 10Licence reads qwen-research 30
nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-GGUF-NInfer90,46734Oct. 5Apache 2.0 31
unsloth/embeddinggemma-2-GGUF71,914272Oct. 6Apache 2.0 32
peasantsmith/GLM-5.3-Flash-Maya-GGUF71,73432Oct. 7MIT 33
AtomicChat/Qwen-Image-2.1-Turbo-GGUF61,81923Oct. 9Licence reads other 34
FastVideo/FastVideo-FastH3-Trim-Comfy26,85927Oct. 5Licence reads other 35
ggml-org/embeddinggemma-2-GGUF25,44148Oct. 6Apache 2.0 36
abhishek085/spark-s1-4b-v6-GGUF22,2120Oct. 10Apache 2.0 37
JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF20,64669Oct. 7Apache 2.0 38
mflux-community/qwen-image-edit-2511-mflux-q420,0290Oct. 10Apache 2.0 39
ConwayResearch/Underdog-Saluki-27B-1.0 is the single largest arrival of the week and the one a builder should open first. It is Qwen3.8-27B cut to an IQ2 mix of 7.89 GB for stock llama.cpp, with tool calling preserved. 27 The card reports 88 against 84 for the full model on its own tool-calling bench, 42 parallel calls against 35, and 96% average retention across nine benchmarks. 27 An optional 928 MB projector adds image input. Those are the authors' numbers on the authors' bench, and a 27B model at two bits is where a card's claims and your workload part company.
LiquidAI/d1-3B and its GGUF build arrive from a lab that already sells an on-device line. The tags describe decision and classification work with calibration and a System One posture, and the licence is the LFM 1.0 agreement, linked from the card as its own document. 40 The GGUF drew 16,778 downloads in six days against 11,866 for the base repository, which is the pattern the whole week keeps showing.
Embedding work produced the other cluster. unsloth/embeddinggemma-2-GGUF and ggml-org/embeddinggemma-2-GGUF together drew 97,355 downloads in six days, off a base repository that has itself reached 115,763. 3241 That base is dated September 14 and sits in neither of this week's earlier snapshots, so it carries no ratio here while its GGUF conversions do.
JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF continues the Mellum line this channel covered in June, now as a 64-expert reasoning model with GGUF builds for llama.cpp, Ollama and LM Studio. 38
The Qwen-Image-2.1-Turbo release on October 9 shows where image-model traffic goes. The official repository holds 6,787 downloads and 609 likes. 42 Repackagings of it hold 331,019 across three repositories, with ComfyUI-loadable files and one abliterated text encoder. 283034 The same split appeared with Qwen-Image-2.1 itself in September. Take the licence from the parent, whatever a repackaging declares.

The rest of the screen

Twenty-four more repositories cleared the ratio above twenty thousand downloads, and eleven of them belong to a single account's 2024 output. 12
RepositoryOct. 6 → Oct. 12ChangeLicense
babybirdprd/Experiment26-7B-passthrough-10slice-GGUF909 → 182,637200.9xno licence tag
cortecs/Llama-3.3-70B-Instruct-FP8-Dynamic8,191 → 83,69610.2xother, linking the Meta Llama 3 licence 43
Jackrong/MLX-Qwopus3.5-9B-v3-4bit690 → 80,453116.6xApache 2.0 44
RichardErkhov/MaziyarPanahi_-_Topxtral-4x7B-v0.1-gguf848 → 74,78188.2xno licence tag
RichardErkhov/SeaLLMs_-_SeaLLMs-v3-7B-Chat-gguf774 → 65,41884.5xno licence tag
openchat/openchat-3.5-12101,792 → 58,67832.7xApache 2.0 45
urchade/gliner_large-v2872 → 58,16366.7xApache 2.0 46
Serkan007/moondream2-multiformat464 → 56,014120.7xApache 2.0, the card's only content 47
RichardErkhov/ibm-granite_-_granite-20b-code-instruct-gguf1,293 → 54,61642.2xno licence tag
RichardErkhov/anthracite-org_-_magnum-12b-v2-gguf226 → 51,869229.5xno licence tag
RichardErkhov/ibm-granite_-_granite-20b-code-base-gguf2,079 → 46,80122.5xno licence tag
RichardErkhov/rbiswasfc_-_Qwen2-1.5B-Instruct-xLAM-fc-gguf582 → 42,98673.9xno licence tag
RichardErkhov/Jiqing_-_tiny_random_falcon-gguf217 → 42,955197.9xno licence tag
RichardErkhov/BeaverAI_-_NeMoist-21B-v1a-gguf240 → 42,255176.1xno licence tag
thesven/Mistral-7B-Instruct-v0.3-GGUF501 → 41,24682.3xApache 2.0 48
zetasepic/Qwen2.5-72B-Instruct-abliterated-GGUF1,360 → 40,22429.6xLicence reads other 49
mradermacher/Llama-3.3-70B-Instruct-abliterated-GGUF3,260 → 36,52311.2xLicence reads llama3.3 50
RichardErkhov/fblgit_-_juanako-7b-UNA-gguf212 → 35,545167.7xno licence tag
JackKozmo29/depguard-jev-style-1.5b823 → 30,78837.4xApache 2.0 24
JackKozmo29/mailguard-jev-style-1.5b1,701 → 30,61418.0xApache 2.0 25
JackKozmo29/codeguard-jev-style-1.5b441 → 29,26466.4xApache 2.0 26
RichardErkhov/Sao10K_-_L3-8B-Stheno-v3.1-gguf426 → 27,44464.4xno licence tag
RichardErkhov/avemio-digital_-_mistral_7b_Sft_pretrain_6000_steps_merge-gguf319 → 24,62377.2xno licence tag
unsloth/llava-1.5-7b-hf-bnb-4bit1,247 → 23,08018.5xLicence reads llama2 51
Every reading in that table comes from the same two snapshots as the rows above. 12
The eleven repositories from one account deserve a longer look. They are GGUF conversions of other people's models, uploaded in June and August 2024, and not one of them carries a Hub licence tag. 1 Together they stood at 7,216 downloads on October 6 and 509,293 at the cutoff, an aggregate of 70.6x. 2 Eleven repositories moving six hundredfold between two readings six days apart is a fact about the counter, which counts file transfers and records nothing about who made them.
The band below holds many more like them. Between 5,000 and 20,000 downloads sit 82 repositories, most of them quantisations, abliterations and runtime repackagings of models already named above.

Before building

  1. Run each candidate on a task from your own product, with the pass-fail threshold fixed before the first run rather than after it.
  2. Measure peak memory, throughput and hosting cost on the machine you intend to ship on. The AWQ VRAM figure, the DGX Spark tokens per second and the Q5_K_M footprint are all the authors' measurements on their hardware.
  3. Read the repository licence and the base model's licence, and follow the stricter one. Four rows this week turn on that: the Gemma 4 NVFP4 build declares Apache 2.0 while its own licence link opens Google's terms, the Qwen3.5 hybrid inherits two parents' terms, the Qwen-Image-2.1-Turbo repackagings declare permissions their parent does not, and HunyuanVideo 1.5 leaves the licence field unresolved.
  4. Treat one counter reading as a level. Direction appears when two dated readings of the same repository are compared, and this issue prints both for every row so you can recompute them.

References

  1. 1
    Hub-wide model snapshot, Oct. 6, 2026

    huggingface.co

  2. 2
    Hub-wide model snapshot, Oct. 12, 2026

    huggingface.co

  3. 3
    Alennnndy/Qwen3.6-27B-AWQ-4bit model card

    huggingface.co

  4. 4
    Qwen3.6-27B model card

    huggingface.co

  5. 5
    Gemma 4 31B IT QAT NVFP4 Blackwell model card

    huggingface.co

  6. 6
    Gemma 4 licence

    ai.google.dev

  7. 7
    Empero Qwen3.8-4B-Distill model card

    huggingface.co

  8. 8
    bleysg hybrid model card

    huggingface.co

  9. 9
    Qwen3.5-122B-A10B-FP8 repository

    huggingface.co

  10. 10
    sevclassifier model card

    huggingface.co

  11. 11
    OneBrain Rerank v1 model card

    huggingface.co

  12. 12
    bge-reranker-v2-m3 model card

    huggingface.co

  13. 13
    CPH-Community-7B model card

    huggingface.co

  14. 14
    mindview-t2i-turbo model card

    huggingface.co

  15. 15
    VoiceTut-TTS GGUF model card

    huggingface.co

  16. 16
    QwenCleo-ASR GGUF model card

    huggingface.co

  17. 17
    hubert-large-korean model card

    huggingface.co

  18. 18
    RynnBrain 1.1-2B model card

    huggingface.co

  19. 19
    RynnBrain 1.1 technical report

    arxiv.org

  20. 20
    imajev-9b-GGUF model card

    huggingface.co

  21. 21
    olmOCR-2 MLX 8-bit model card

    huggingface.co

  22. 22
    allenai/olmOCR-2-7B-1025 model card

    huggingface.co

  23. 23
    HunyuanVideo-1.5 Diffusers model card

    huggingface.co

  24. 24
    depguard model card

    huggingface.co

  25. 25
    mailguard model card

    huggingface.co

  26. 26
    codeguard model card

    huggingface.co

  27. 27
    Underdog-Saluki-27B-1.0 model card

    huggingface.co

  28. 28
    AtomicChat Qwen-Image-2.1-Turbo-Uncensored model card

    huggingface.co

  29. 29
    kindlingai glm-5.3 exlr8 model card

    huggingface.co

  30. 30
    unsloth Qwen-Image-2.1-Turbo GGUF model card

    huggingface.co

  31. 31
    nerkyor Qwen3.8-27B NInfer model card

    huggingface.co

  32. 32
    unsloth embeddinggemma-2 GGUF model card

    huggingface.co

  33. 33
    GLM-5.3-Flash-Maya GGUF model card

    huggingface.co

  34. 34
    AtomicChat Qwen-Image-2.1-Turbo GGUF model card

    huggingface.co

  35. 35
    FastVideo FastH3 Trim Comfy model card

    huggingface.co

  36. 36
    ggml-org embeddinggemma-2 GGUF model card

    huggingface.co

  37. 37
    spark-s1-4b-v6 GGUF model card

    huggingface.co

  38. 38
    Mellum2.1 Thinking GGUF model card

    huggingface.co

  39. 39
    mflux qwen-image-edit q4 model card

    huggingface.co

  40. 40
    LiquidAI/d1-3B-GGUF model card

    huggingface.co

  41. 41
    google/embeddinggemma-2 repository

    huggingface.co

  42. 42
    Qwen/Qwen-Image-2.1-Turbo repository

    huggingface.co

  43. 43
    Llama-3.3-70B-Instruct-FP8-Dynamic model card

    huggingface.co

  44. 44
    MLX-Qwopus3.5-9B-v3-4bit model card

    huggingface.co

  45. 45
    openchat-3.5-1210 model card

    huggingface.co

  46. 46
    GLiNER-large-v2 model card

    huggingface.co

  47. 47
    moondream2-multiformat card

    huggingface.co

  48. 48
    Mistral-7B-Instruct-v0.3-GGUF model card

    huggingface.co

  49. 49
    Qwen2.5-72B-Instruct-abliterated-GGUF model card

    huggingface.co

  50. 50
    Llama-3.3-70B-Instruct-abliterated-GGUF model card

    huggingface.co

  51. 51
    llava-1.5-7b-hf-bnb-4bit model card

    huggingface.co

Hugging Face Surging Models
Hugging Face Surging Models@NeoDrop Official

This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.

Related content

  • LLM
  • Image generation
  • Audio
  • Multimodal
  • Repositories that did not exist when the window opened
  • The rest of the screen
  • Before building