xClean Tools

AI Models — Daily Top 5

UPDATED 2026-09-06 01:03 PDT

2026-09-06 · SUNDAY · 01:03 PDT

  1. 1 Qwen/Qwen3.8-27B 6,024,467 DOWNLOADS · 14,056 LIKES Qwen3.8-27B is the Qwen team's 27B-parameter vision-language model, released under Apache 2.0 as post-trained weights in Hugging Face Transformers format. The card lists compatibility with Transformers, vLLM, SGLang and TokenSpeed, with a hosted API offered separately through Qwen Cloud. It is the most liked model in today's set at 14,056 likes and about 6.0 million downloads.
  2. 2 XHToken/Spark-X2.5-4B 4,755 DOWNLOADS · 563 LIKES Spark-X2.5-4B is a 4B-parameter open text-generation model from XHToken, shipped for conversational use with custom modelling code alongside the safetensors weights. Its companion project presents the Spark-X2.5 series as an attempt to push agentic capability into models that run on device. It is the newest entry here, with about 4,755 downloads and 563 likes.
  3. 3 facebook/mms-300m 12,464 DOWNLOADS · 263 LIKES facebook/mms-300m is the 300 million parameter checkpoint from Meta's Massively Multilingual Speech project, pretrained with the wav2vec2 self-supervised objective on roughly 500,000 hours of audio across more than 1,400 languages. It is a base model intended to be fine-tuned for a downstream speech task and expects 16 kHz input. It shows about 12,464 downloads and 263 likes.
  4. 4 distilbert/distilbert-base-uncased 7,054,316 DOWNLOADS · 1,156 LIKES DistilBERT base uncased is a distilled, smaller version of BERT base that does not distinguish letter case. It is published for masked language modelling and serves mainly as a starting point for fine-tuning, which keeps it among the Hub's most used encoders at roughly 7.1 million downloads. The checkpoint carries 1,156 likes.
  5. 5 openai/clip-vit-base-patch32 20,579,479 DOWNLOADS · 1,210 LIKES CLIP ViT-B/32 pairs images with free-form text labels, which lets it classify pictures into categories it was never explicitly trained on. OpenAI researchers built it to study what makes computer vision models robust, and it remains a default choice for zero-shot image classification and embeddings. At about 20.6 million downloads it is the most downloaded model in today's set, with 1,210 likes.

2026-09-05 · SATURDAY · 13:05 PDT

  1. 1 IFM/K2-Horizon-MoVA-36B-A4B 1,333 DOWNLOADS · 170 LIKES IFM released the sparse member of its K2-Horizon family, a mixture-of-experts model that holds 36B parameters but runs about 4B per token, paired with mixture-of-values attention. The card claims frontier-class agentic and reasoning results at that active-parameter budget. Only the final checkpoint is out; intermediate checkpoints, data and training code are promised later.
  2. 2 Jackrong/Qwopus3.8-27B-Flash-GGUF 10,680 DOWNLOADS · 110 LIKES A llama.cpp-ready GGUF build of Qwopus3.8-27B-Flash, a 27B multimodal model that takes image and text input, packaged for local inference. The author has posted a warning that a flaw in second-stage reinforcement learning makes the model emit wrong indentation in some Python programs, and plans to retrain that stage. It has been pulled about 10.7k times.
  3. 3 lightx2v/Minimax-h3-Turbo 1,185,646 DOWNLOADS · 842 LIKES LightX2V's turbo distillation of MiniMax-H3 does text-to-video, image-to-video and reference-to-video in a handful of sampling steps. At roughly 1.19M downloads it is by far the most pulled model on the board, and third-party INT8 ComfyUI repacks that fuse it into a single file are trending alongside it. A hosted studio demo is available.
  4. 4 XHToken/Spark-X2.5-1.7B 2,301 DOWNLOADS · 90 LIKES XHToken published Spark-X2.5-1.7B, a small conversational text-generation model shipped in safetensors with custom modeling code. The 1.7B size puts it in range of local and on-device deployment. It has about 2.3k downloads and 90 likes so far.
  5. 5 nvidia/Qwen3.8-Flash-Next-NVFP4 1,129 DOWNLOADS · 83 LIKES NVIDIA published an NVFP4 four-bit quantization of Alibaba's Qwen3.8-Flash-Next, produced with its Model Optimizer toolkit. The base is a causal language model with a vision encoder, hybrid Gated DeltaNet and sparse attention, a mixture-of-experts stack and n-gram embeddings. The quantized weights cut the memory needed to serve it.

2026-09-04 · FRIDAY · 13:05 PDT

  1. 1 DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU 2,022 DOWNLOADS · 153 LIKES DavidAU published an uncensored 27B merge of Qwen3.8 tagged for image-text-to-text use, shipping safetensors weights through transformers. The author bills it as the first of several planned releases, with GGUF conversions and more than ten further variants in a companion repository. It is the only model on today's trending list, with 153 likes and about 2,000 downloads.

2026-09-03 · THURSDAY · 13:06 PDT

  1. 1 OpenVDN/vdn-minimax-h3 0 DOWNLOADS · 126 LIKES OpenVDN released VDN-Minimax-H3, which applies Video DeltaNet's hybrid attention to the MiniMax H3 video generator for what the card describes as near-lossless quality at lower cost. On eight B200 GPUs the authors report generating video faster than it plays back. Weights, code and license ship together as a text-to-video diffusers model.
  2. 2 DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF 39,646 DOWNLOADS · 122 LIKES DavidAU published MTP GGUF quants of a Qwen3.8 27B fine-tune the card calls TURBO, tuned to cut thinking tokens by half or more while keeping output detail. The card claims scores above 735 on ARC-C and 880 on ARC-E in 8-bit, and above 718 ARC-C at 4-bit. It has drawn nearly 40,000 downloads.
  3. 3 AngelSlim/Hy4-preview-GGUF 97,192 DOWNLOADS · 89 LIKES AngelSlim published three GGUF builds of Tencent's Hy4-preview, from a standard 4-bit Q4_K_M at 435 GiB down to about 214 GiB using the UD-IQ1_M and MIX_STQ1_0 strategies. The two smaller builds roughly halve the file size for local runs of a very large model. It is the most downloaded model in today's pool at about 97,000 downloads.

2026-09-02 · WEDNESDAY · 01:04 PDT

  1. 1 unsloth/Qwen3.8-27B-GGUF 9,354,057 DOWNLOADS · 3,355 LIKES Unsloth published GGUF builds of Qwen3.8-27B using its Dynamic 3.0 quantization, which the model card says holds accuracy better than other leading quants. The card also notes developer-role support so the model works inside agentic tools such as Codex, and improved parsing of nested tool-call objects. The repo has 9.4M downloads and 3,355 likes.
  2. 2 Lightricks/LTX-2.5 1,232,274 DOWNLOADS · 2,488 LIKES Lightricks released LTX-2.5, a video generation model published in single-file diffusion format. Its repository tags span image-to-video, text-to-video and video-to-video generation as well as audio-to-video and video-to-audio conversion. It has drawn 1.2M downloads and 2,488 likes.
  3. 3 unsloth/GLM-5.3-Flash-GGUF 63,718 DOWNLOADS · 327 LIKES Unsloth released GGUF quantizations of GLM-5.3-Flash, a bilingual English and Chinese text generation model. The card directs users to a llama.cpp pull request or the Unsloth desktop app to run the files, and demonstrates a 1-bit build of the model running in that desktop UI. The repo has 63,718 downloads and 327 likes.
  4. 4 MiniMaxAI/MiniMax-H3 5,532,597 DOWNLOADS · 4,772 LIKES MiniMax published MiniMax-H3, a video generation model covering text-to-video, image-to-video and video-to-video along with joint text-to-audio-video output. The card points to hosted APIs on the company's global and China platforms, its web apps, and official prompt-writing skills on GitHub. It has 5.5M downloads and 4,772 likes.
  5. 5 Qwen/Qwen3.8-Flash-Next-FP8 130,451 DOWNLOADS · 181 LIKES Qwen published FP8-quantized weights for the post-trained Qwen3.8-Flash-Next in Hugging Face Transformers format. The card describes fine-grained FP8 quantization with a block size of 128 and reports metrics nearly identical to the unquantized model, with compatibility across Transformers, vLLM and SGLang. It has 130,451 downloads and 181 likes.

2026-09-01 · TUESDAY · 13:05 PDT

  1. 1 google/timesfm-3.0-pytorch 0 DOWNLOADS · 188 LIKES Google Research published the official PyTorch weights and configurations for TimesFM 3.0, its pretrained time-series forecasting foundation model. The card describes a stacked mixing transformer with variate attention and releases the model under a non-commercial license. The repository has drawn 188 likes with no downloads recorded yet.
  2. 2 google-bert/bert-base-uncased 69,651,344 DOWNLOADS · 2,831 LIKES BERT base uncased remains one of the most downloaded checkpoints on Hugging Face, logging 69,651,344 downloads and 2,831 likes. The English masked language model comes from the original 2018 BERT paper and does not distinguish case. It is distributed for PyTorch, TensorFlow, JAX, ONNX, Core ML and Rust.
  3. 3 sentence-transformers/all-MiniLM-L6-v2 255,143,740 DOWNLOADS · 5,350 LIKES all-MiniLM-L6-v2 leads today's set with 255,143,740 downloads and 5,350 likes. The sentence-transformers model maps sentences and paragraphs into a 384-dimensional dense vector space for clustering and semantic search, and ships in PyTorch, ONNX, OpenVINO and Rust builds.
  4. 4 Momoking/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4 0 DOWNLOADS · 98 LIKES This repository re-quantizes an uncensored Qwen3-VL-32B text encoder for MiniMax-H3 video generation into mixed-precision NVFP4, cutting it to 15.7 GB so it fits on a single 16 GB card. The build is packaged for ComfyUI and has drawn 98 likes with no downloads recorded yet.
  5. 5 openai-community/gpt2 14,502,665 DOWNLOADS · 3,502 LIKES GPT-2 is still in heavy use, with 14,502,665 downloads and 3,502 likes. The English causal language model was introduced by OpenAI in the paper Language Models are Unsupervised Multitask Learners and is distributed for PyTorch, TensorFlow, JAX, TFLite, ONNX and Rust.

2026-08-31 · MONDAY · 13:06 PDT

  1. 1 deepseek-ai/DeepSeek-V4-Flash-Vision-Exp 0 DOWNLOADS · 304 LIKES DeepSeek published DeepSeek-V4-Flash-Vision-Exp, the first experimental multimodal model in its V4 family, and it tops today's Hugging Face trending set with 304 likes. The release adds visual modules to the DeepSeek-V4-Flash architecture and continues training to unlock image understanding. DeepSeek says it improves substantially on multimodal agent tasks over DeepSeek-V4-Flash-0731.
  2. 2 Kijai/MiniMax-H3-experimental 0 DOWNLOADS · 374 LIKES Kijai posted experimental repacks of MiniMax-H3 for ComfyUI, which have drawn 374 likes, the most in today's set. They include a w4a8 format pairing 4-bit weights with int8 convrot activations, and an int8 convrot VAE the author says cuts decode time by about a third; both need ComfyUI 0.31.0. A four-step distilled video checkpoint and an experimental reference LoRA also ship.
  3. 3 incoai/GLM-5.3-Flash-DFlash2 7,322 DOWNLOADS · 90 LIKES Inco AI released DFlash 2, a draft model for speculative decoding with GLM-5.3-Flash, which has logged 7,322 downloads. The repository is not a standalone language model: it runs inside a speculative decoding server and drafts tokens for the target model to verify. DFlash 2 is a block-diffusion drafter that predicts a whole block of tokens in a single pass.
  4. 4 DavidAU/Qwen3.8-27B-Cold-Fable-Fusion-GAIN-V1.1-732-Heretic-Uncensored-stage1 17 DOWNLOADS · 91 LIKES DavidAU published a first-stage Cold Fusion build on Qwen3.8-27B that has collected 91 likes on 17 downloads. The repository is tagged as an uncensored fine-tune using the Heretic method and is registered for image-text-to-text use, so it takes both images and text. Its model card is currently empty.
  5. 5 ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF 18,665 DOWNLOADS · 88 LIKES ISTA-DASLab released non-uniform GGUF quantizations of Qwen3.8-27B, the most downloaded entry in today's set at 18,665. The files are produced with the lab's GSQ and RCO methods, which mix precision across the model rather than applying a single bit width, and they ship with a vision projector for multimodal use.