xClean Tools

AI Models — Daily Top 5

UPDATED 2026-11-02 01:15 PST

2026-11-02 · MONDAY · 01:15 PST

  1. 1 microsoft/Mage-VL-Cowork 752 DOWNLOADS · 126 LIKES Microsoft released Mage-VL-Cowork, a 4B computer-use model trained to decide when to click a GUI and when to write code or shell commands for the same task. Its card reports 66.0 percent on OSWorld-Verified and describes a full training recipe learned from evolved harnesses, starting from the intermediate Mage-VL checkpoint. It is Apache-2.0 licensed.
  2. 2 FusionCow/Sulphur-3 0 DOWNLOADS · 136 LIKES Sulphur-3 is a finetune of MiniMax-H3 whose repository ships a 40 GB bf16 checkpoint labelled Ref2VA-hybrid beside int8 and W4A8 builds, a corrector checkpoint, a rank-256 LoRA and a ComfyUI bundle. Access to the repository is gated and the model card is frontmatter only, so no evaluation or usage detail is published. It has 136 likes and no recorded downloads.

2026-11-01 · SUNDAY · 13:12 PST

  1. 1 OBLITERATUS/DeepSeek-V4.1-Flash-1017-obliterated 520 DOWNLOADS · 130 LIKES A refusal-ablated build of DeepSeek-V4.1-Flash-1017 was released as BF16 safetensors under the MIT license, produced with the OBLITERATUS toolkit using the failspy method. Its model card reports full refusals falling from 81.6 percent to 4.3 percent across 838 evaluation prompts, at a KL divergence of 0.1305 nats from the original. The same card records harmful assistance rising from 10.4 percent to 78.3 percent.

2026-10-31 · SATURDAY · 01:13 PDT

  1. 1 Qwen/Qwen3.8-Flash-Next 4,137,335 DOWNLOADS · 6,661 LIKES Qwen3.8-Flash-Next is an open-weight experimental preview of the architecture that will underpin Qwen4. The model carries 125B parameters with 6B activated per token plus 51B of n-gram embeddings, and pairs Gated DeltaNet with Qwen Sparse Attention, which selects at micro-block level to cut long-context latency. Weights ship in Transformers format.
  2. 2 Edge0/Audio8-TTS-Realistic-0.6b 813 DOWNLOADS · 127 LIKES Audio8-TTS-Realistic-0.6b is a 0.6B-parameter text-to-speech model fine-tuned from Edge0's TTS preview for more natural speech across 23 languages. It supports zero-shot voice cloning and keeps the parameter count of the earlier release. The Apache-2.0 weights come with a demo space and high-throughput serving instructions.

2026-10-30 · FRIDAY · 13:15 PDT

  1. 1 speridlabs/idol 221 DOWNLOADS · 130 LIKES Idol turns one picture of an object into a 3D Gaussian splat, generating up to 1,048,576 Gaussians including the sides the picture never showed and writing them as a standard 3DGS .ply file. It removes the background and frames the object itself, uses 7.2GB of VRAM on a consumer GPU and can finish in under two seconds. Sperid Labs released it under Apache-2.0 with an SDK and a demo Space.
  2. 2 maya-research/maya2-native 1,493 DOWNLOADS · 125 LIKES Maya Research released Maya 2 Native, an Apache-2.0 streaming text-to-speech model covering 11 Indian languages and English, with 20 built-in voices and zero-shot cloning from 3 to 10 seconds of reference audio. A cloned voice carries across languages, so a Hindi reference can speak Tamil. The card places it second for Hindi on VoiceArena's real-time leaderboard as of October 27.

2026-10-29 · THURSDAY · 13:18 PDT

  1. 1 mistralai/Mistral-Large-4-1T-A52B 278 DOWNLOADS · 1,178 LIKES Mistral published Mistral Large 4, a 1 trillion-parameter natively multimodal mixture-of-experts model with 49 billion active parameters per token, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers. It takes text and images, carries a 1M-token context and exposes three reasoning efforts, and the weights ship in FP8 with an embedded speculative-decoding draft head.
  2. 2 Diffuspace/Clover-35B-A3B-Preview 1,030 DOWNLOADS · 136 LIKES DiffuSpace released its first model, Clover-35B-A3B-Preview, a diffusion language model derived from Qwen3.6-35B-A3B with about 3 billion active parameters per token and a 128K context. It keeps the base model's hybrid Gated DeltaNet and full-attention MoE backbone, adds bidirectional block attention, and writes text by resolving confident masked positions block by block.
  3. 3 LiquidAI/LFM3-Audio-2B 1,794 DOWNLOADS · 132 LIKES Liquid AI published LFM3-Audio-2B, a compact audio model meant to run locally on laptops and edge devices. It takes text or audio and answers with text, speech or both, offering English voice chat plus transcription, synthesis and translation across 18 languages. The repo bundles its own audio tokenizer and modeling code.
  4. 4 hi3d-official/Hi3Dv1.0 0 DOWNLOADS · 129 LIKES Hi3D v1.0 refines a coarse mesh into high-resolution 3D geometry, conditioned on a reference image of the same object and built on the Sparc3D sparse representation. The project report describes training on roughly two million curated 3D assets and three refinement passes reaching a 1536-cubed sparse extraction grid. It was released on October 23.

2026-10-28 · WEDNESDAY · 13:18 PDT

  1. 1 openbmb/MiniCPM-OCR-1.0 791 DOWNLOADS · 238 LIKES MiniCPM-OCR 1.0 is a 1.3B Apache-2.0 model and the first OCR-specific entry in the MiniCPM-V/o family. It targets text-centric document understanding across text, tables, formulas, charts, seals and structured extraction, and reports 93.68 on Real5-OmniDocBench, a benchmark for real-world document parsing. Separate DSpark checkpoints add decoding acceleration.
  2. 2 google/embeddinggemma-2 282,912 DOWNLOADS · 2,991 LIKES EmbeddingGemma 2 is an open multimodal embedding model from Google DeepMind that maps text and code, images, video and audio, and combinations of them, into a single 768-dimensional vector space. It totals 740M parameters, pairing a 270M text model with modular 170M vision and 300M audio encoders, under an Apache-2.0 license. Google designed it to run on consumer hardware such as phones and laptops.
  3. 3 LiquidAI/LFM3-Omni-2B 714 DOWNLOADS · 130 LIKES LFM3-Omni-2B is Liquid AI's first omni-modal model in the LFM3 family, a single 2.2B-parameter network that understands text, images, video and audio and answers in text or speech. Liquid says one checkpoint replaces a separate speech recognition model, vision-language model, LLM and text-to-speech model. For videos with sound it processes picture and soundtrack together.
  4. 4 nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B 338 DOWNLOADS · 127 LIKES Nemotron 3.5 Transcribe is NVIDIA's open-weight offline speech model, transcribing 40 languages from recordings up to one hour with a 65,536-token context. The Mamba-2 and Transformer hybrid mixture of experts totals 34B parameters with about 4.3B active, and also handles speech translation, speaker-attributed transcription, timestamps and audio question answering. It ships in BF16, FP8 and NVFP4.
  5. 5 LiquidAI/LFM3-VL-2B 1,596 DOWNLOADS · 128 LIKES LFM3-VL-2B is an on-device vision-language model that pairs the LFM3 language model with a new in-house vision encoder and takes text, images and video as input. It is the first LFM vision model to accept video, reasons step by step before answering, and targets spatial and embodied tasks such as locating objects for robotics. Liquid does not recommend it for coding or multi-turn conversation.

2026-10-27 · TUESDAY · 13:18 PDT

  1. 1 baidu/Unlimited-OCR-2 513 DOWNLOADS · 216 LIKES Baidu released Unlimited OCR Works 2, a vision-language document parser the model card bills as moving from long-horizon to unlimited-horizon parsing through a technique it calls SR-SWA. The MIT-licensed transformers checkpoint arrived on October 27 together with MarathonDocBench, a new benchmark for ultra-long document parsing, plus ModelScope and Baidu Cloud mirrors.
  2. 2 ccore/Qwen3.5-4B-3D-Coder 3,544 DOWNLOADS · 157 LIKES ccore published an early preview of Qwen3.5-4B-3D-Coder, an independent fine-tune of Qwen/Qwen3.5-4B aimed at general coding and at generating interactive Three.js scenes. The Q4_K_M and Q8_0 GGUF builds were refreshed on October 26 to fix a layer issue; the author says training finishes on October 28, with full FP16 weights, the dataset and an evaluation report to follow.
  3. 3 LiconStudio/Minimax-H3-VBVR 4,701 DOWNLOADS · 125 LIKES Licon VBVR is an image-to-video LoRA for MiniMax-H3, trained on VBVR-Pro-SFT-Video data rebuilt around that base model. The Apache-2.0 adapter targets object consistency and action logic, cutting unintended duplicate objects, clipping and interpenetration during motion, and the card ships side-by-side clips against the unmodified base.