xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-09-12 · SATURDAY · 13:05 PDT

  1. 1 Alissonerdx/Minimax-H3-ComfyUI 10,390 DOWNLOADS · 120 LIKES This repository packages LoRAs for MiniMax H3 video generation to run inside ComfyUI, shipping an LMS LoRA together with a ready-made workflow. The weights were trained and tested mainly against the ref2va base checkpoint, with fl2va listed as a workable but less-tested base. It has about 10,400 downloads and 120 likes.
  2. 2 Agnes-AI/Agnes-3.0-Flash 34 DOWNLOADS · 111 LIKES Agnes AI published an open-weight preview checkpoint of Agnes 3.0 Flash, a 33B multimodal model with a 262,144-token context window. The repository states it is a distinct, earlier checkpoint from the production API model, which uses a different configuration and a 1M-token context. It has drawn 111 likes on 34 downloads so far.

2026-09-11 · FRIDAY · 13:06 PDT

  1. 1 Qwen/Qwen-Drive-1.0-4B 3,271 DOWNLOADS · 167 LIKES Qwen released Qwen-Drive-1.0-4B, an autonomous-driving model that keeps the architecture of the pretrained Qwen3.5 vision-language model and folds 3D perception, visual question answering and motion planning into one framework. Weights and configs sit on Hugging Face, with code, demo data and documentation in a separate QwenLM repository.
  2. 2 orcarouter/Qwen3.8-27B-Uncensored-GGUF 320,537 DOWNLOADS · 893 LIKES An uncensored GGUF build of Qwen3.8-27B is the most downloaded model on the board, at 320,537 pulls. The repository packages abliterated weights, with refusal behaviour stripped out, in llama.cpp's format for local image-text-to-text use. A second abliterated GGUF conversion of the same base model is also trending today.
  3. 3 tencent/AuK 554 DOWNLOADS · 92 LIKES Tencent's Hunyuan group published AuK, an open-source foundation model for speech generation and editing. One model covers zero-shot text to speech, voice cloning, speech editing and speech enhancement, and the release ships with a paper, a GitHub repository and a Hugging Face demo space.
  4. 4 TokenRhythm/NeoHorse-1-4B 6,109 DOWNLOADS · 1,084 LIKES NeoHorse-1-4B is a 4B text model from TokenRhythm built through agentic post-training with a routing harness, which its technical report frames as a step toward recursive self-improvement. It runs on a Qwen3.5 text backbone and is tagged for agentic use, tool calling, coding and reasoning. With 1,084 likes it is the most liked model on the board.
  5. 5 Edge0/Edge0-35B-A3B-preview 951 DOWNLOADS · 98 LIKES Edge0 published a preview of Edge0-35B-A3B, a 35B-class sparse mixture-of-experts model the authors say runs in phone-class memory: 3 GiB of active memory at 4-bit, around 15 tokens per second. It uses a Qwen3.5 MoE architecture with a pre-router and SSD offload, and ships in MLX format.

2026-09-10 · THURSDAY · 13:05 PDT

  1. 1 XHToken/Spark-X2.5-4B-GGUF 73,777 DOWNLOADS · 133 LIKES Spark-X2.5-4B-GGUF is a BF16 GGUF conversion of Spark-X2.5-4B, a compact general-purpose model for conversation, writing, translation, reasoning, coding, tool use and agentic workflows. The base model uses a hybrid attention architecture, supports a native context length of up to 1M tokens and covers more than 200 languages. The build targets llama.cpp, Ollama and LM Studio.
  2. 2 phasefield-audio/Irodori-TTS-v4.1-Anime 0 DOWNLOADS · 99 LIKES Irodori-TTS-v4.1-Anime is a Japanese text-to-speech model fine-tuned from Aratako's Irodori-TTS-v4.1-Small on anime-style speech data. Because the base model's annotation pipeline is not publicly documented, the fine-tuning data was annotated independently, so caption conditioning and emoji controls may behave differently from the base. Quantized checkpoints ship alongside the full-precision one.
  3. 3 JonathanColetti/Qwen3.8-27B-Uncensored-GGUF 2,812,743 DOWNLOADS · 1,047 LIKES This repository publishes GGUF quantizations of Qwen3.8-27B with the multi-token prediction head retained and verified, alongside imatrix and speculative-decoding support. The uploader states that refusal behavior is substantially reduced rather than eliminated, and that capabilities, training data and architecture are otherwise unchanged. It is the most-downloaded new entry in the day's pool at 2.8 million.
  4. 4 facebook/sam3 1,840,407 DOWNLOADS · 3,050 LIKES SAM 3 is Meta's Segment Anything checkpoint for mask generation, distributed in safetensors for the transformers library and carrying a video variant in the same repository. It is the most-liked entry in the day's Hugging Face pool at 3,050 likes, with roughly 1.84 million downloads.
  5. 5 krea/Krea-2-Turbo 70,893 DOWNLOADS · 1,136 LIKES Krea-2-Turbo is a text-to-image model from Krea, published in safetensors for the diffusers library. It has drawn 1,136 likes against roughly 70,900 downloads, one of the highest like-to-download ratios in the day's pool.

2026-09-09 · WEDNESDAY · 13:06 PDT

  1. 1 nex-agi/Nex-N2.5-Pro 0 DOWNLOADS · 164 LIKES Nex-AGI's Nex-N2.5-Pro sits on the Hugging Face trending board with 164 likes and no downloads yet, because the weights are still listed as coming soon. Pro is the middle tier of the Nex-N2.5 family, alongside mini and Max, which Nex-AGI positions for long-horizon agentic tasks in real-world environments. The card says the line builds on the multimodal foundations of Nex-N2.

2026-09-08 · TUESDAY · 13:07 PDT

  1. 1 XHToken/Spark-X2.5-4B 10,661 DOWNLOADS · 835 LIKES Spark-X2.5-4B is a 4-billion-parameter text generation model from XHToken, released in the Spark-X2.5 series with a custom architecture that runs through remote code. It is the most-liked new model on Hugging Face today at 835 likes, with about 10,700 downloads.
  2. 2 Qwen/Qwen3.8-27B 6,712,160 DOWNLOADS · 14,376 LIKES Qwen3.8-27B carries the post-trained weights of Alibaba's 27B multimodal model in Hugging Face Transformers format, taking image and text input and producing text. The repository lists compatibility with Transformers, vLLM and SGLang under an Apache 2.0 license. It has drawn about 6.7 million downloads and 14,400 likes.
  3. 3 facebook/mms-300m 12,388 DOWNLOADS · 294 LIKES Meta's Massively Multilingual Speech model at 300 million parameters is back on the trending list. It was pretrained with the Wav2Vec2 self-supervised objective on roughly 500,000 hours of speech across more than 1,400 languages, and the card says it is meant to be fine-tuned for a downstream task rather than used as is.
  4. 4 distilbert/distilbert-base-uncased 7,138,152 DOWNLOADS · 1,183 LIKES DistilBERT base uncased remains one of the most-downloaded models on Hugging Face, at about 7.1 million pulls. The distilled version of BERT base is case-insensitive and serves masked language modelling, still acting as a small default encoder years after its release.
  5. 5 openai/clip-vit-base-patch32 20,702,763 DOWNLOADS · 1,240 LIKES OpenAI's CLIP ViT-B/32 is the most-downloaded model in today's set at about 20.7 million pulls. It pairs a vision transformer with a text encoder for zero-shot image classification, and its card notes it was built to study robustness and generalization in vision rather than for deployed use.

2026-09-07 · MONDAY · 13:03 PDT

  1. 1 openbmb/MiniCPM5-2B 13 DOWNLOADS · 169 LIKES OpenBMB released MiniCPM5-2B, the second model in the MiniCPM5 series after MiniCPM5-1B. It is a dense 2B transformer that scales up the same training recipe for on-device, local and resource-constrained deployment, and the release claims 2B-class open-source state of the art. It is the most-liked new model on Hugging Face today.
  2. 2 WarmBloodAban/Minimax-h3_Singularity 26,731 DOWNLOADS · 137 LIKES Minimax-h3_Singularity is a fine-tuned fusion of MiniMax-H3 checkpoints built for multimodal video generation in ComfyUI. A single model covers text-to-video, image-to-video, reference-to-video and video-to-video workflows. With roughly 26,700 downloads it leads the day's new models on that measure.
  3. 3 Viggle/Viggle-Animate 0 DOWNLOADS · 102 LIKES Viggle-Animate swaps the character in a video using a single repainted frame, leaving motion, camera and timing untouched. The user prepares that one frame in an image editor and the video stage propagates the change through the clip. It is built on MiniMax-H3 with distillation, and ships with a demo Space.
  4. 4 IFM/K2-Horizon-7B 2,087 DOWNLOADS · 86 LIKES K2-Horizon-7B is the medium dense member of the K2-Horizon family, a decoder-only model with a native 512K-token context window carried from midtraining onward. It is evaluated across agentic, coding, long-context and reasoning benchmarks. Intermediate checkpoints are released as well, so capability changes can be studied across training.

2026-09-06 · SUNDAY · 13:02 PDT

  1. 1 dealignai/GLM-5.3-CYBERSECURITY-FP8 15,648 DOWNLOADS · 126 LIKES A community fine-tune of GLM-5.3 with refusal behavior reduced for security tasks has drawn about 15,600 downloads and 126 likes on Hugging Face. The uploader ships it in native FP8 for Hopper GPUs and frames it as a domain-specific release for offensive security rather than a general uncensored model. Releases like this keep raising questions about safety training being stripped and redistributed.
  2. 2 microsoft/VibeVoice-ASR-Streaming-7B 889 DOWNLOADS · 99 LIKES Microsoft published VibeVoice-ASR-Streaming-7B, a streaming speech recognition model that transcribes both who is speaking and what they said. It supports customized hotwords and 10 languages, and ships alongside code and a live playground in the VibeVoice project. Early uptake is 889 downloads and 99 likes.
  3. 3 MATLOWAI/minimax-h3-fused-turbo-int8-convrot 53,074 DOWNLOADS · 97 LIKES A single 21 GB ComfyUI file packages MiniMax-H3 text and image-to-video plus reference-to-video generation in four steps, with a distilled turbo and a motion-smoothing LoRA folded into the INT8 weights. The merge lets one file serve both first/last-frame and reference conditioning instead of a separate model per mode. It has about 53,000 downloads and 97 likes.