xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-09-28 · MONDAY · 13:08 PDT

  1. 1 ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF 58,021 DOWNLOADS · 138 LIKES ukisai has published compact mixed-precision GGUF builds of its Swift 1.5 Qwen3.8-27B model for llama.cpp. The quantization reuses the per-tensor bit allocations from ISTA-DASLab's GSQ-RCO release with Swift-specific refinement, and the underlying 1.5 model extends Swift 1.0 with post-training aimed at long-horizon agentic work.
  2. 2 Alissonerdx/BFS-Best-Face-Swap 150,020 DOWNLOADS · 1,005 LIKES BFS is a collection of LoRA adapters for face, head and body replacement in image generation. Separate releases target Qwen Image 2.1, Qwen Image Edit, Flux 2 Klein and Krea 2, each with its own base model, input order and prompt contract, alongside experimental LTX-2 workflows.

2026-09-27 · SUNDAY · 13:10 PDT

  1. 1 SupersonicLabs/Julia-1 274 DOWNLOADS · 176 LIKES Supersonic Labs released Julia 1, a 144.3M-parameter decision model that turns a state, a question and candidate answers into one typed decision, covering classification, routing, ordered scores and Boolean calls. Measured on 2026-09-24 it hit 73.15% on a 2,000-case typed-decision suite and 71.50% macro accuracy across 52 MASSIVE locales.
  2. 2 SupraLabs/Supra2-IMG 515 DOWNLOADS · 148 LIKES Supra Labs' Supra2-IMG is a 104M-parameter diffusion transformer that generates 256x256 images from text, trained from scratch for 10 epochs on 5.6M captioned images from FLUX-Reason-6M. It pairs a frozen Flan-T5-Base encoder with SD-VAE-FT-MSE, took about 9 hours on a single H100, and ships under Apache-2.0 with an inference script.

2026-09-26 · SATURDAY · 13:09 PDT

  1. 1 DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF 1,603,479 DOWNLOADS · 1,198 LIKES GGUF quants of DavidAU's Qwen3.8-27B TURBO fine-tune, which the author reports as the first to pass 730 on ARC-C and 880 on ARC-E at 8-bit, and 718 on ARC-C at 4-bit. TURBO refers to cutting thinking tokens by half to as much as a tenth while holding output detail. The repo has drawn 1.6 million downloads.
  2. 2 fastino/GLiNER2.5-Decide 14,772 DOWNLOADS · 184 LIKES GLiNER2.5-Decide is a 340M-parameter English classification model that accepts any label set at call time - intent, routing, sentiment, priority, policy or multi-label tags - in a single forward pass. It needs no prompt template and generates no tokens, loading locally through AutoExtractor. One call can score several heads at once.
  3. 3 moondream/parakeet-redux 1,636 DOWNLOADS · 159 LIKES Moondream released a 1.58-bit rebuild of NVIDIA's parakeet-tdt-0.6b-v3 speech recognizer in which every encoder weight is -1, 0 or +1, on the same architecture and tokenizer. It fits in 178 MB and runs at 113 times real time on eight x86 CPU cores, which the team measures as 2.5 times the fastest other Parakeet runtime. It stays within 0.3 WER of the original on English and beats it on the 25-language FLEURS set.
  4. 4 orcarouter/OrcaSAQ-2-27B 1,182 DOWNLOADS · 153 LIKES OrcaSAQ-2-27B compresses Qwen3.8-27B from a 54 GB BF16 checkpoint to 12.3 GB using OrcaRouter's proprietary 3-bit mixed-precision quantization. The publisher reports a 0.02 percent perplexity change, 93.2 percent top-1 agreement with the original and a 262K context window. It is aimed at long-horizon agents, coding, tool use and stateful execution.

2026-09-25 · FRIDAY · 13:11 PDT

  1. 1 StarDoc-AI/TeleOCR 32,056 DOWNLOADS · 300 LIKES An Apache-2.0 document parsing model built on Qwen2.5-VL that handles clean digital files and photographs of pages in one system. The card points to a GitHub repository and carries a leaderboard badge claiming a top score on OmniDocBench v1.6. It has been downloaded about 32,000 times.
  2. 2 Viggle/Qwen-Image-2.1-viggle-turbo 47,873 DOWNLOADS · 221 LIKES A few-step student of Qwen-Image-2.1 that Viggle trained with distribution matching distillation. It generates images and instruction-driven edits from one to three reference images in six transformer passes instead of 40 and without classifier-free guidance, which the card puts at roughly five times faster end to end than the base model.
  3. 3 apple/LensVLM-9B 812 DOWNLOADS · 208 LIKES Apple's 9B vision-language model reads text as compressed images, then uses learned tools to expand only the pages it actually needs back to full resolution. It is built on Qwen3.5-9B and ships alongside a paper and code on selective context expansion for long visual text.
  4. 4 Contrastive-LM/CLM-v0.1-8B 13 DOWNLOADS · 196 LIKES A state head and an action head sit on top of a frozen Qwen3-8B encoder and are trained with a bidirectional InfoNCE objective, so the model scores how well an action fits a state instead of generating text. It was pre-trained on about 60 million question-answer pairs and mid-trained on 30 million synthetic hard negatives, and is pitched as a verifier and reranker for agents.

2026-09-24 · THURSDAY · 13:10 PDT

  1. 1 Lightricks/LTX-2.5 1,637,601 DOWNLOADS · 4,987 LIKES LTX-2.5 is Lightricks' video generation model, published in single-file diffusion format. Its repository tags cover text-to-video, image-to-video and video-to-video generation as well as joint audio-video output. The repo is now gated, granting access automatically to signed-in users who accept its terms.
  2. 2 Edge0/Audio8-ASR-Infinite 347 DOWNLOADS · 274 LIKES Edge0 released Audio8-ASR-Infinite, a natively streaming speech recognition model for Chinese and English that decodes 12.5 times per second with a configurable 240 to 560 millisecond transcription delay. A rolling KV cache holds memory and latency constant, which the card says allows unlimited-length transcription in 24/7 operation.
  3. 3 nvidia/Nemotron-3-Diarization 4,282 DOWNLOADS · 258 LIKES NVIDIA released Nemotron 3 Diarization on September 23, an open-weight speaker diarization model that works out who spoke when for up to eight speakers. A single checkpoint serves both streaming and offline inference, with input buffer latency as low as 80 milliseconds, and chunked inference removes any limit on audio length.
  4. 4 MiniMaxAI/MiniMax-H3 3,640,535 DOWNLOADS · 5,650 LIKES MiniMax-H3 is MiniMax's video generation model, covering text-to-video, image-to-video and video-to-video along with synchronized audio and video output. The card points to hosted APIs on the company's global and China platforms and to official prompt-writing skills on GitHub. It has 3.6 million downloads and 5,650 likes.
  5. 5 akhilaaa3/Jev-Omni 0 DOWNLOADS · 191 LIKES Jev-Omni is a multimodal decision classifier built on Gemma 4 12B IT with a 30,000-question fine-tuning run. Given a question and a set of options over text, images, audio or video, it returns a probability for each option instead of a generated explanation. The card reports 87.57 percent on DecisionBench Medium.

2026-09-23 · WEDNESDAY · 13:11 PDT

  1. 1 convaiinnovations/laya-multilingual 0 DOWNLOADS · 218 LIKES Convai Innovations released Laya Multilingual, a 322M-parameter checkpoint in its non-autoregressive System 1 decision family covering more than 100 languages. Built on an mmBERT-base encoder with a 1,024-token context, it takes a state and typed questions and returns typed answers with probabilities in one forward pass, generating no text. The company positions it as the checkpoint for anything that is not English.
  2. 2 pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF 74,291 DOWNLOADS · 187 LIKES pottokao published an abliterated text encoder for Qwen-Image-2.1 in three formats in one repo: a 5.0 GB Q4_K_M GGUF paired with a separate mmproj file, a 9.3 GB FP8 safetensors build for NVIDIA GPUs, and a 17.5 GB bf16 build for Mac and other non-CUDA setups. A companion ComfyUI node fixes GGUF loading of the Qwen3-VL vision tower.
  3. 3 inclusionAI/Ming-Image-0.1-Design 0 DOWNLOADS · 175 LIKES inclusionAI released Ming-Image-0.1-Design, a 6B text-to-image model aimed at UI, infographics, posters and other text-rich layouts. It generates complete visual compositions and supports RGBA output with transparent backgrounds, triggered by prepending one of the recommended phrases to the prompt. The MIT-licensed weights ship with a companion inference repository and a hosted demo.
  4. 4 AikidoSec/altar-1 951 DOWNLOADS · 159 LIKES Aikido Security published Altar-1, a 504B-parameter prune of GLM-5.3 that strips 34 percent of the 753B original's experts and quantizes to INT4, landing at 328 GB to serve on four H200 GPUs under vLLM. The pruning used REAP, which scores each expert by router-weighted activation, calibrated on cybersecurity traces, coding, tool calling, reasoning and multilingual Wikipedia.

2026-09-22 · TUESDAY · 13:10 PDT

  1. 1 XiaomiMiMo/MiMo-V2.6-Pro-RL 985 DOWNLOADS · 374 LIKES MiMo-V2.6-Pro-RL is the flagship checkpoint of Xiaomi's MiMo-V2.6 series, a natively omnimodal model that takes text, image, video and audio in a 1M-token context. Xiaomi trained it in one mixed reinforcement learning run across coding, agent, visual and cybersecurity tasks, using asynchronous GRPO on batches of 1,568 prompts by 16 rollouts with an agentic grader that ranks rollouts within each group.
  2. 2 XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B 804 DOWNLOADS · 324 LIKES Xiaomi also released MiMo-V2.6-Distill-Qwen-9B, a 9B agentic model built by supervised fine-tuning of Qwen3.5-9B on MiMo-generated data for code, general agent work, visual coding and cybersecurity. The mixture holds 77.4B tokens, 27.2B of them loss-bearing. The checkpoint scores 44.6 on SWE Pro against 32.0 for the base model and is offered as a starting point for open agentic RL research.
  3. 3 zai-org/GLM-5.3-Flash 3,547,021 DOWNLOADS · 2,527 LIKES GLM-5.3-Flash, Z.ai's first natively multimodal GLM-5 model, is trending again with about 3.5 million downloads in the past month. It activates 18B of 320B parameters, combines sparse and linear attention to cut long-context serving cost, adds Manifold-Constrained Hyper-Connections and trains on a 30T-token multimodal corpus. Z.ai says it beats GLM-5.2 at one-tenth the price.
  4. 4 XGENlabs/XGEN-JING 2 DOWNLOADS · 167 LIKES XGEN-JING is an egocentric interactive experience model built on MiniMax-H3 that generates first-person video and audio from actions, reference images and observation history. Keyboard input drives camera movement while text guides object interaction and dialogue. The release ships a four-step bidirectional model, inference code and prompt skills; the causal model and technical report are still to come.
  5. 5 DeepMostInnovations/sales-conversion-model-reinf-learning 432 DOWNLOADS · 177 LIKES DeepMost Innovations' sales conversion model is a PPO-trained reinforcement learning policy that tracks how each turn of a sales conversation shifts the probability of closing. It ships through the deepmost Python package, optionally pairing with a local LLM for dynamic metrics. The model card reports 96.7 percent conversion prediction accuracy and 85 ms inference, and cites the SalesRLAgent paper.