xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-09-21 · MONDAY · 13:10 PDT

  1. 1 Altworld/Hemmingway-1 834 DOWNLOADS · 344 LIKES Altworld's Hemmingway-1 is a 27B open-weights model built on Qwen3.8-27B for everyday writing such as messages, emails and awkward notes, returning the text itself rather than options and commentary. On the team's own blind CommunicationBench of 80 requests it ranked first, ahead of Fable 5.1 and GPT-6 Astra, and it placed third on the public EQ-Bench 4. It is Apache-2.0 with a 262,144-token context.
  2. 2 netease-youdao/Confucius4-R2T2 1,864 DOWNLOADS · 220 LIKES NetEase Youdao's Confucius4-R2T2 is a streaming speech-recognition model built on Qwen3-ASR-1.7B that commits transcript text as it arrives and never revises earlier words. Decoding chunks can be set from 80 ms to 2 s, and the card reports 200 to 600 ms average latency with accuracy close to offline recognition. It is tuned for Chinese and English and ships with a vLLM backend.
  3. 3 yandex/AliceAI-Foundation-80B-A3B-Base 676 DOWNLOADS · 172 LIKES Yandex released AliceAI-Foundation-80B-A3B-Base, a base language model trained from scratch that activates 3B of its 80B parameters per token and reads up to 262,144 tokens of context. Its 48 layers pair KDA and gated-attention blocks with a 512-expert mixture of experts. Yandex says it matches larger open models on math and coding and is especially strong on Russian facts; the weights are Apache-2.0.
  4. 4 bespokelabs/Bespoke-Nimble-9B 1,138 DOWNLOADS · 144 LIKES Bespoke Labs' Nimble is a LoRA adapter of about 165 MiB for Qwen3.5-9B that answers context-grounded multiple-choice, yes-or-no and rubric-score questions by scoring the allowed answer tokens directly instead of generating reasoning or free text. Given a context and a typed schema, it returns each field's answer with probabilities; fields take up to 26 choices and prompts are capped at 2,048 tokens.
  5. 5 jinaai/jina-ocr-v1 2,615 DOWNLOADS · 137 LIKES Jina AI's jina-ocr-v1 is a document-parsing model built on DeepSeek-OCR, pairing its DeepEncoder vision tower with a 3B mixture-of-experts decoder that activates about 570M parameters per token, plus a FastMTP speculative-decoding head. The card reports 91.14 on OmniDocBench v1.6, 83.4 on olmOCR-Bench, 7.4 points above DeepSeek-OCR, and 2.57 pages per second on one A100. It is licensed CC-BY-NC-4.0.

2026-09-20 · SUNDAY · 13:04 PDT

  1. 1 Qwen/Qwen-Image-2.1 183 DOWNLOADS · 593 LIKES Alibaba's Qwen team open-sourced Qwen-Image-2.1, a unified model for text-to-image generation and image editing. Its visual generation component uses 7B parameters across 32 single-stream DiT layers, which the team presents as a balance of output quality, inference efficiency and versatility. The repository has collected 593 likes since posting.
  2. 2 Cactus-Compute/needle3 36,096 DOWNLOADS · 126 LIKES Cactus Compute released Needle3, a foundation model for phones, wearables, robots and microcontrollers that ships as a single file of 8 to 29 MB. The model trades general chat ability for on-device tool calling and extraction, and its card claims it beats models ten times its size on mobile tool calls. It has been downloaded about 36,000 times.
  3. 3 Accio-Lab/occamy-1.0 1,659 DOWNLOADS · 154 LIKES Accio Lab published Occamy-1.0, a 35B mixture-of-experts model aimed at agentic tool use and working alongside people. The release ships GGUF, FP8 and NVFP4 builds plus an experimental multi-token prediction head, and the lab describes the weights as an open Pareto frontier for that size class. The repo has 154 likes and about 1,700 downloads.

2026-09-19 · SATURDAY · 13:09 PDT

  1. 1 TokenRhythm/NeoHorse-1-9B 11,692 DOWNLOADS · 934 LIKES NeoHorse-1-9B is a text-only model TokenRhythm post-trained from Qwen3.5-9B for agent harnesses, tool use, coding and instruction following. It was built with a routing harness that assigns tasks to a pool of models and feeds the outcomes into the next training mix, which the team calls an early step toward recursive self-improvement. It reports a 69.04 average on ten benchmarks, 3.44 points above Qwen3.5-9B.
  2. 2 fastino/gliner2.5-multi-v1 200,044 DOWNLOADS · 217 LIKES GLiNER2.5 Multi is Fastino's multilingual information-extraction model, a 287M-parameter checkpoint on mDeBERTa-v3-base that handles entities, text classification, structured records, relations and span attributes from one schema. Its boundary architecture pairs span starts and ends instead of using a fixed span-width grid, and it runs locally on CPU, CUDA or Apple's MPS. It has passed 200,000 downloads.
  3. 3 cyjin-yl/Qwen3.8-27B-Uncensored-Cyber-agentic-imatrix-GGUF 28,383 DOWNLOADS · 151 LIKES This repository holds GGUF quantizations of philbert440's Qwen3.8-27B-Uncensored-Cyber, a de-refused Qwen3.8-27B for cyber and offensive-security questions, calibrated with an importance matrix built from real agentic-coding sessions instead of generic prose. The aim is to keep tool-call JSON, chat-template tokens and long verbatim strings such as file paths intact; IQ4_XS files ship with and without the MTP head.
  4. 4 pyannote/speaker-diarization-community-1 5,300,239 DOWNLOADS · 1,892 LIKES Community-1 is pyannote's open speaker-diarization pipeline: it takes mono 16 kHz audio, downmixing or resampling other input, and outputs who spoke when. It improves speaker assignment and counting over the 3.1 pipeline, adds an exclusive mode that lines up more easily with transcripts, and runs offline; it is CC-BY-4.0 licensed, with downloads gated behind a contact form. It has passed 5.3 million downloads.

2026-09-18 · FRIDAY · 13:12 PDT

  1. 1 internlm/Atria-Dawn-Preview 711 DOWNLOADS · 162 LIKES Shanghai AI Laboratory released Atria Dawn Preview, an agentic model built on the 744B-parameter mixture-of-experts GLM-5.2 and aimed at deep research, software building, document delivery and authorized security work. Its model card reports the best BrowseComp (92.5) and CyberGym (86.5) scores among the compared models, though it trails Claude Opus 5 on SWE-bench Pro. It is MIT-licensed with a 256K context.
  2. 2 audnai/penclaw-GLM-5.3-abliterated 565 DOWNLOADS · 161 LIKES audnai published penclaw-GLM-5.3-abliterated, an uncensored build of Z.ai's GLM-5.3 made with abliteration, a technique that strips out a model's refusal behavior. The Transformers-format weights and model card sit behind an access request, so the repository says little publicly beyond its tags. It has drawn 161 likes and 565 downloads.
  3. 3 Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4 0 DOWNLOADS · 139 LIKES This release adds real-audio tooling to the YuE2-3B music model: an audio-to-semantic-token encoder that YuE2 does not ship, plus a LoRA for its non-autoregressive decoder branch. Together they let users tokenize their own recordings, fine-tune YuE2 on an artist and generate new songs or covers, with every stage fitting on a 24 GB GPU.
  4. 4 deepseek-ai/DeepSeek-R1 746,600 DOWNLOADS · 14,261 LIKES DeepSeek-R1 is DeepSeek's first-generation reasoning model, a 671B-parameter mixture-of-experts with 37B active parameters and a 128K context, trained with cold-start data followed by large-scale reinforcement learning. Its model card reports performance comparable to OpenAI-o1 on math, code and reasoning tasks. The MIT-licensed repository has about 14,300 likes and 747,000 downloads.
  5. 5 AlexWortega/openjev 0 DOWNLOADS · 124 LIKES openjev turns Qwen3.5-4B into a single cross-encoder that reads a premise and a hypothesis and labels the pair as entailment, contradiction or neutral. The author applies that one primitive, without task-specific training, to reranking answers, grading them against a reference, guarding content and playing Doom zero-shot from text or pixels. The MIT-licensed release also covers a larger Qwen3.5-35B-A3B variant.

2026-09-17 · THURSDAY · 13:07 PDT

  1. 1 ukisai/Swift-Qwen3.8-27b 3,221 DOWNLOADS · 374 LIKES UkisAI published Swift-Qwen3.8-27B, a reasoning-efficient derivative of Qwen3.8-27B. The team reports it spends 58.3 percent fewer thinking tokens for under 1 percent performance loss, giving a 1.95x speed-up on several tasks, by targeting reasoning-marker tokens it identifies as triggering overthinking. A GGUF build ships alongside the main weights.
  2. 2 harshatheg/Qwen-2.5-1B-RLCD 0 DOWNLOADS · 248 LIKES Qwen-2.5-1B-RLCD ships as part of a parallel constrained decoding engine for Apple Silicon, aimed at high-throughput structured information extraction. The MLX-based release targets JSON and classification output and comes with a side-by-side demo on Hugging Face Spaces. It has drawn 248 likes without registering downloads yet.
  3. 3 Qwen/Qwen3.8-Flash-Next 706,052 DOWNLOADS · 5,358 LIKES The Qwen team published Qwen3.8-Flash-Next, a post-trained multimodal model released in Hugging Face Transformers format. The weights are described as compatible with Transformers, vLLM and SGLang, with a managed version offered through the Qwen Cloud API. The repository has passed 700,000 downloads and 5,300 likes.
  4. 4 dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 32,011 DOWNLOADS · 263 LIKES Dealign AI published an abliterated FP8 build of DeepSeek-V4.1-Flash, with the original model's safety guardrails removed at the weight level. The uploader states that vision, reasoning, multi-token prediction and the 1M-token context of the base model are preserved. The repository has logged about 32,000 downloads.
  5. 5 meta-llama/Llama-3.1-8B-Instruct 5,887,953 DOWNLOADS · 7,692 LIKES Meta's Llama-3.1-8B-Instruct is back on the Hugging Face trending list, with close to 5.9 million downloads and more than 7,600 likes. The instruction-tuned 8B text model remains among the most downloaded open-weight checkpoints on the hub.

2026-09-14 · MONDAY · 01:10 PDT

  1. 1 meta-llama/Llama-3.1-8B-Instruct 5,620,539 DOWNLOADS · 7,587 LIKES Meta's Llama-3.1-8B-Instruct is the 8-billion-parameter instruction-tuned text-generation model in the Llama 3.1 family, distributed in safetensors and PyTorch formats for the Transformers library. It is the most downloaded model on today's list, with about 5.62 million downloads and 7,587 likes.
  2. 2 deepseek-ai/DeepSeek-R1 648,837 DOWNLOADS · 14,230 LIKES DeepSeek-R1 is one of DeepSeek's first-generation reasoning models, introduced with DeepSeek-R1-Zero, which was trained by large-scale reinforcement learning without supervised fine-tuning as a preliminary step. The card says R1-Zero gained strong reasoning behaviors from RL but struggled with issues such as endless repetition. Its 14,230 likes are the most on today's list, alongside about 648,800 downloads.
  3. 3 dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 2,254 DOWNLOADS · 130 LIKES dealignai's release applies what the uploader calls permanent weight-level abliteration to DeepSeek-V4.1-Flash, stripping its safety guardrails while, it says, preserving MMLU capability, vision, reasoning and multi-token prediction. It ships in native FP8 with a 1-million-token context, tool support and maximum reasoning as the default. It has about 2,250 downloads and 130 likes.
  4. 4 thesysdev/OUI-1 1,447 DOWNLOADS · 105 LIKES OUI-1 from thesysdev is billed as the first diffusion model built for generative UI, a fine-tune of Google's DiffusionGemma 26B-A4B-it that writes user interface screens in openui-lang, the declarative language behind OpenUI. It scores 71.7% on the Generative UI Benchmark, 5.5 times its base model, with 4B active parameters. It has about 1,450 downloads and 105 likes.
  5. 5 huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF 2,682,158 DOWNLOADS · 711 LIKES huihui-ai's GGUF release is an uncensored build of Qwen3.8-27B made with abliteration, which the card calls a crude proof of concept for removing refusals without TransformerLens. Its latest update adds a Huihui-Qwen3.8-27B-abliterated-GSQ-RCO series. With about 2.68 million downloads and 711 likes, it is the second most downloaded model on today's list.

2026-09-13 · SUNDAY · 13:11 PDT

  1. 1 deepseek-ai/DeepSeek-V4.1-Flash 244,457 DOWNLOADS · 2,179 LIKES DeepSeek-V4.1-Flash from DeepSeek is a multimodal mixture-of-experts model with 552B backbone parameters and support for contexts of up to one million tokens. It processes images and text natively and uses a causal encoder-decoder design built on a 40-layer transformer, with a technical report centered on KV cache compression. The model has about 2,180 likes and 244,500 downloads.
  2. 2 nex-agi/Nex-N2.5-mini 3,970 DOWNLOADS · 754 LIKES Nex-N2.5-mini is the smallest of Nex-AGI's three Nex-N2.5 agentic models, alongside Pro and Max, built for long-horizon tasks in real-world environments. Like Pro, it builds on the multimodal foundations of Nex-N2, with focused improvements in computer use. Its tags point to a Qwen3.5 MoE architecture, and it has about 3,970 downloads and 754 likes.
  3. 3 m-a-p/YuE2-3B 3,707 DOWNLOADS · 389 LIKES YuE2-3B from m-a-p is an open music generation model that turns lyrics and a style prompt into a complete song with vocals and accompaniment, whose melody and chords can then be reshaped through an editable score. The team claims song quality competitive with Suno v5 and v6 and state-of-the-art results on WildSongBench. It has about 3,710 downloads and 389 likes.
  4. 4 openbmb/MiniCPM5-2B-GGUF 99,716 DOWNLOADS · 214 LIKES OpenBMB's MiniCPM5-2B-GGUF is the GGUF build of MiniCPM5-2B, the second model in the MiniCPM5 series after MiniCPM5-1B. The base is a dense 2B transformer made for on-device, local and resource-constrained deployment, which OpenBMB says reaches 2B-class open-source state of the art. The GGUF repository has about 99,700 downloads and 214 likes.
  5. 5 HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF 2,157,412 DOWNLOADS · 1,145 LIKES HauhauCS publishes an aggressive uncensored variant of Qwen3.8-27B in GGUF format, tuned to answer directly without refusal behavior, and reports 0 refusals across 465 test prompts. The build adds the publisher's FastMTP speculative decoding, claimed to deliver up to 3.02 times the document token generation speed of a non-MTP build. It has about 2.16 million downloads and 1,145 likes.