Search FoxAI
AI Models 183
- Qwen3.8-Flash-Next-NVFP4↗
primitive-ai
Qwen3.8-Flash-Next is 360 GB in BF16 and needs two data-center GPUs in FP8. This build serves it on one 96 GB Blackwell : 88.8 GiB of VRAM, the 51B n-gram table in host RAM, and no patched runtime. Two flags are load-bearing on a single GPU. `--distributed-executor-backend mp`: the default single-GPU executor never starts the n-gram offload worker, and the first forward waits on it forever — the server looks healthy and hangs. `VLLM_PLE_OFFLOAD_READY_TIMEOUT=1800`: the worker loads a 95 GB table before serving and the 600 s default can expire first. Host needs about 100 GB of free RAM. The reasoning parser and tool-call parser in this block are validated on this exact checkpoint: thinking lands in `reasoning` with no markup in `content`, and tool calls arrive as structured `tool_calls` with valid JSON arguments and `finish_reason: tool_calls`. If you have seen other NVFP4 cards warn that the automatic MoE backend picks `FLASHINFER_TRTLLM` and corrupts output: on this image on an RTX PRO 6000 the automatic choice is `FLASHINFER_CUTLASS`, printed at boot, and every number on this card went through it. Measured on the mixed sibling (same expert kernel), `--moe-backend marlin` decodes 10% faster single-stream (93.2 vs 84.6 tok/s) and 1–3% slower at concurrency 32, with TTFT up 12%; tool-calling under Marlin 79.5 (n=1). Marlin does not combine with MTP on this image: the BF16 MTP draft inherits the backend and refuses it, and the draft-scoped override in `--speculative-config` is ignored by Model Runner V2 until vllm-project/vllm#54788 lands. Choose MTP for interactive use, Marlin only without speculation. On vLLM nightly builds from early September the SM120 selector picks Marlin by default anyway (vllm#54788); the flag matters on this image. Two more flags are in the block since 2026-09-06: without them the image's defaults do not boot on a 96 GB card. A field report found it first and we reproduced both failures on this box the same day. Without MTP the default `--max-num-seqs` (1024) exceeds the Mamba state blocks left at 0.92 utilization (598 here) and CUDA graph capture aborts with `Engine core initialization failed`; with MTP at the native 262,1
- Qwen-Image-Layered↗
阿里巴巴 · Qwen
  🤗 HuggingFace    |   🤖 ModelScope    |    📑 Research Paper    |    📑 Blog    |    🤗 Demo    We are excited to introduce **Qwen-Image-Layered**, a model capable of decomposing an image into multiple RGBA layers. This layered representation unlocks **inherent editability**: each layer can be independently manipulated without affecting other content. Meanwhile, such a layered representation naturally supports **high-fidelity elementary operations**-such as resizing, reposition, and recoloring. By physically isolating semantic or structural components into distinct layers, our approach enables high-fidelity and consistent editing. Given an image, Qwen-Image-Layered can decompose it into several RGBA layers:
- Qwen-Image-2.1↗
阿里巴巴 · Qwen
🤖 ModelScope | 🤗 HuggingFace | 📑 Blog | 🖥️ Demo | 🫨 Discord | 💬 WeChat We are excited to open-source **Qwen-Image-2.1**, a unified text-to-image generation and image editing model in the Qwen family. With just **7B parameters in its visual generation component** (32 Single-Stream DiT layers), Qwen-Image-2.1 balances generation quality, inference efficiency, and versatility. - **Compact and Efficient** — A lightweight architecture with mixed-granularity attention and prefix KV cache reuse delivers strong image quality at low computational cost. - **Native Transparency, Unified Creation and Editing** — Generate regular or transparent (RGBA) images from text, edit transparent layers, and extract subjects from photographs—all in one model. - **Versatile Editing** — Support up to **10 reference images**, specify local edits via circles, painted annotations, or separate masks, and preserve identity for people and products. - **Realistic Textures and Refined Aesthetics** — Improved typography, portrait lighting, and fine details for more visually compelling results.
- Qwen-Image-Layered-GGUF↗
unsloth
> [!NOTE] > This is a GGUF quantized version of Qwen-Image-Layered. > unsloth/Qwen-Image-Layered-GGUF uses Unsloth Dynamic 2.0 methodology for SOTA performance. Important layers are upcasted to higher precision.   🤗 HuggingFace    |   🤖 ModelScope    |    📑 Research Paper    |    📑 Blog    |    🤗 Demo    We are excited to introduce **Qwen-Image-Layered**, a model capable of decomposing an image into multiple RGBA layers. This layered representation unlocks **inherent editability**: each layer can be independently manipulated without affecting other content. Meanwhile, such a layered representation naturally supports **high-fidelity elementary operations**-such as resizing, reposition, and recoloring. By physically isolating semantic or structural components into distinct layers, our approach enables high-fidelity and consistent editing.
- yovoice-model↗
leemysw
This repository has GGUF model packages for yovoice that are not in the official audio.cpp model repository. Each package runs with the audio.cpp inference engine. The same files are mirrored on Hugging Face and ModelScope. You don't need to download these files by hand. yovoice downloads them from **Settings › Models** (or `yovoice models download `) and checks the size and SHA256 of each file. - 103 built-in voices: 100 Chinese voices and 3 English voices (`af_maple`, `af_sol`, `bf_vale`). No reference audio or voice cloning. - Languages: Chinese, American English, British English. Experimental. - Includes word-level Chinese pronunciation tables made with the official Misaki frontend, so inference runs without Python. - Bundled resources and licenses (copies are inside the GGUF under `licenses/`): eSpeak-ng data (GPL-3.0-or-later), Misaki (Apache-2.0), jieba (MIT), pypinyin (MIT).
- Wan2.2-I2V-A14B-Diffusers↗
Wan-AI
💜 Wan    ｜    🖥️ GitHub    |   🤗 Hugging Face    |   🤖 ModelScope    |    📑 Technical Report    |    📑 Blog    |   💬 WeChat Group    |    📖 Discord    We are excited to introduce **Wan2.2**, a major upgrade to our foundational video models. With **Wan2.2**, we have focused on incorporating the following innovations: - 👍 **Effective MoE Architecture**: Wan2.2 introduces a Mixture-of-Experts (MoE) architecture into video diffusion models. By separating the denoising process cross timesteps with specialized powerful expert models, this enlarges the overall model capacity while maintaining the same computational cost.
- gemma-scope-2-1b-it↗
Google
- modelscope-damo-text2video-pruned-weights↗
kabachuha
- Qwen-Image-Edit-2511-FP8↗
drbaph
This is the fp8_e4m3fn quantized version of Qwen-Image-Edit-2511 for lower memory usage and faster inference (compatible with ComfyUI) approx. 20GB VRAM 💜 Qwen Chat    |   🤗 Hugging Face    |   🤖 ModelScope    |    📑 Tech Report    |    📑 Blog    🖥️ Demo    |   💬 WeChat (微信)    |   🫨 Discord   |    Github   
- Wan2.1-Fun-14B-Control↗
alibaba-pai
V1.0: | 名称 | 存储空间 | Hugging Face | Model Scope | 描述 | |--|--|--|--|--| | Wan2.1-Fun-1.3B-InP | 19.0 GB | 🤗Link | 😄Link | Wan2.1-Fun-1.3B文图生视频权重，以多分辨率训练，支持首尾图预测。 | | Wan2.1-Fun-14B-InP | 47.0 GB | 🤗Link | 😄Link | Wan2.1-Fun-14B文图生视频权重，以多分辨率训练，支持首尾图预测。 | | Wan2.1-Fun-1.3B-Control | 19.0 GB | 🤗Link | 😄Link| Wan2.1-Fun-1.3B视频控制权重，支持不同的控制条件，如Canny、Depth、Pose、MLSD等，同时支持使用轨迹控制。支持多分辨率（512，768，1024）的视频预测，支持多分辨率（512，768，1024）的视频预测，以81帧、每秒16帧进行训练，支持多语言预测 | | Wan2.1-Fun-14B-Control | 47.0 GB | 🤗Link | 😄Link| Wan2.1-Fun-14B视频控制权重，支持不同的控制条件，如Canny、Depth、Pose、MLSD等，同时支持使用轨迹控制。支持多分辨率（512，768，1024）的视频预测，支持多分辨率（512，768，1024）的视频预测，以81帧、每秒16帧进行训练，支持多语言预测 | Windows 的详细信息： - 操作系统 Windows 10 - python: python3.10 & python3.11 - pytorch: torch2.2.0 - CUDA: 11.8 & 12.1 - CUDNN: 8+ - GPU： Nvidia-3060 12G & Nvidia-3090 24G Linux 的详细信息： - 操作系统 Ubuntu 20.04, CentOS - python: python3.10 & python3.11 - pytorch: torch2.2.0 - CUDA: 11.8 & 12.1 - CUDNN: 8+ - GPU：Nvidia-V100 16G & Nvidia-A10 24G & Nvidia-A100 40G & Nvidia-A100 80G