Saved toolsSubmit a tool
FOXAI SEARCH

Search FoxAI

AI Models 2119

  • Qwen3.6-27B-AWQ-4bit↗

    Alennnndy

    This repository contains an optimized, 4-bit quantized version of the original Qwen3.6-27B model by Alibaba Cloud. The quantization was performed using the **AWQ** algorithm via the `llmcompressor` library, drastically reducing the VRAM required for local inference without a perceptible degradation in the original model's accuracy. * **Base Model:** `Qwen/Qwen3.6-27B` * **Algorithm:** AWQ * **Scheme:** W4A16_ASYM (4-bit asymmetric weights, 16-bit activations) * **Ignored Modules:** `lm_head`, `visual`, `vision` (Vision encoder layers are kept in their original precision to maintain multimodal capabilities). * **Weights Format:** Safetensors * **Calibration Dataset:** `HuggingFaceH4/ultrachat_200k` (512 random samples from the `train_sft` split, formatted with the model's chat template. Max sequence length: 1024).

  • FLUX-3-Image-Klein-pup-Uncensored-GGUF↗

    AtomicChat

    BFL's prompt upsampler for FLUX 3 Image Klein (a 9B Qwen3.5 fine-tune) with its refusal direction projected out, as GGUF for llama.cpp, plus the same edit as a LoRA for the stock files. Klein's text encoder without refusals is in FLUX-3-Image-Klein-TE-Uncensored-GGUF . The measurements are public. Built by the team behind Atomic Chat , the app that runs open-weight models locally and offline. As a chat model, refusals fall from 63% to 0% on held-out English prompts. MMLU goes from 74.15% to 73.95%. As Klein's upsampler, under BFL's t2i system prompt, the stock model refuses none of our requests but softens 2 of 35 sensitive ones in its captions; this one keeps them. The image model draws neither of the two in any build , so on Klein's pictures we measured no uncensoring effect; the open pipeline already draws the sensitive categories ( the per-category table ). Against the other uncensored upsamplers on our stand, prithiv's measures the same as this one, and Sakura's still refuses 22% in English. The same rungs and layouts as our stock upsampler GGUFs, with the edit baked in. KLD and top-1 are against the original BF16 on a 30-language neutral text set (87 chunks of 4096 tokens), so they include the edit itself. Refusals are on held-out prompts (81 English harmful requests; 250 XSTest safe prompts), plain chat, thinking off.

  • Qwen3.8-27B-Abliterated-GSQ-Orca-GGUF↗

    ayozhee

    > **Update — September 27, 2026: IQ3_S-MTP is now available!** > > Added Qwen3.8-27B-Abliterated-GSQ-Orca-IQ3_S-MTP.gguf — **12.12 GB / 11.29 GiB**, built with the same selective GSQ–Orca method. Feedback and issue reports are welcome! > > **The detailed results below refer to the original IQ2_S release.** A compact, experimental GGUF with MTP included, built for local inference on smaller GPUs. It combines GSQ's quantized weights with selected abliterated weights from OrcaRouter and was developed and tested on an **RTX 3060 12 GB**. This is my first open-source model release. My goal is to share something useful for local use, with a straightforward account of how it was built and tested. Abliteration—the weight modification used to reduce refusal behavior—comes from OrcaRouter.

  • bge-m3↗

    BAAI

    In this project, we introduce BGE-M3, which is distinguished for its versatility in Multi-Functionality, Multi-Linguality, and Multi-Granularity. - Multi-Functionality: It can simultaneously perform the three common retrieval functionalities of embedding model: dense retrieval, multi-vector retrieval, and sparse retrieval. - Multi-Linguality: It can support more than 100 working languages. - Multi-Granularity: It is able to process inputs of different granularities, spanning from short sentences to long documents of up to 8192 tokens. We recommend to use the following pipeline: hybrid retrieval + re-ranking. - Hybrid retrieval leverages the strengths of various methods, offering higher accuracy and stronger generalization capabilities. A classic example: using both embedding retrieval and the BM25 algorithm. Now, you can try to use BGE-M3, which supports both embedding and sparse retrieval. This allows you to obtain token weights (similar to the BM25) without any additional cost when generate dense embeddings. To use hybrid retrieval, you can refer to Vespa and Milvus. The previous test results were lower because we mistakenly removed the passages that have the same id as the query from the search results. After correcting this mistake, the overall performance of BGE-M3 on MIRACL is higher than the previous results, but the experimental conclusion remains unchanged. The other results are not affected by this mistake. To reproduce the previous lower results, you need to add the `--remove-query` parameter when using `pyserini.search.faiss` or `pyserini.search.lucene` to search the passages.

  • pythia-12b↗

    EleutherAI

    The *Pythia Scaling Suite* is a collection of models developed to facilitate interpretability research (see paper). It contains two sets of eight models of sizes 70M, 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B. For each size, there are two models: one trained on the Pile, and one trained on the Pile after the dataset has been globally deduplicated. All 8 model sizes are trained on the exact same data, in the exact same order. We also provide 154 intermediate checkpoints per model, hosted on Hugging Face as branches. The Pythia model suite was deliberately designed to promote scientific research on large language models, especially interpretability research. Despite not centering downstream performance as a design goal, we find the models match or exceed the performance of similar and same-sized models, such as those in the OPT and GPT-Neo suites. Previously, we released an early version of the Pythia suite to the public. However, we decided to retrain the model suite to address a few hyperparameter discrepancies. This model card lists the changes ; see appendix B in the Pythia paper for further discussion. We found no difference in benchmark performance between the two Pythia versions. The old models are still available, but we suggest the retrained suite if you are just starting to use Pythia. **This is the current release.**

  • Hy-MT2-cmf↗

    infosave

    Tencent's Hy-MT2 translation family — 33 languages, instruction-following translation (terminology, style, placeholders, structured data) — as CMF files for `cortiq`: one memory-mapped file per model with the weights, the vendor tokenizer, the exact chat template and per-tensor hashes. The same file runs on the CPU, on Vulkan (NVIDIA / AMD / Intel), DX12 and Apple Metal; no Python, no PyTorch, no CUDA toolkit. **Requires cortiq 0.7.0 or newer.** Earlier runtimes do not know the `hunyuan_v1_dense` / `hy_v3` architectures (and 0.6.9 crashes on any 4-bit embedding — fixed in the same release). ¹ Twelve 512-token windows of wikitext-2 test, exact attention, the same yardstick every card on this account uses. The bf16-class reference for the 1.8B (`q8_2f`, not published) scores 24.77 — the 4-bit file sits within 10% (27.13 vs 24.77) of it. These are translation models: their perplexity on English encyclopedia prose is not what they were trained for, which is why column ² exists.

  • sam2_hiera_tiny.fb_r896_2pt1↗

    timm

  • deepseek-moe-16b-chat↗

    DeepSeek

    See the Introduction for more details. Avoiding the use of the provided function `apply_chat_template`, you can also interact with our model following the sample template. Note that `messages` should be replaced by your input. **Note:** By default (`add_special_tokens=True`), our tokenizer automatically adds a `bos_token` (` `) before the input text. Additionally, since the system prompt is not compatible with this version of our models, we DO NOT RECOMMEND including the system prompt in your input.

  • Belle-whisper-large-v3-turbo-zh↗

    BELLE-2

    If you find this model helpful, please *like* this model and star us on https://github.com/LianjiaTech/BELLE and https://github.com/shuaijiang/Whisper-Finetune Fine tune whisper-large-v3-turbo-zh to enhance Chinese speech recognition capabilities, Belle-whisper-large-v3-turbo-zh demonstrates a **24-64%** relative improvement in performance to whisper-large-v3-turbo on Chinese ASR benchmarks, including AISHELL1, AISHELL2, WENETSPEECH, and HKUST. Same to Belle-whisper-large-v3-zh-punct, the punctuation marks come from model punc_ct-transformer_cn-en-common-vocab471067-large, and are added to the training datasets.

  • Minimax_h3_latent_Upscaler↗

    LBH-123-AI

    Neural latent-space upscaler for **Minimax H3** video generation. Works directly on Minimax H3's **24-channel VAE latents** to upscale spatial resolution (H×W) while preserving the time dimension. 1. Generate video at low resolution — far fewer latent tokens, much faster. 2. Upscale the latent in-place with this learned upscaler. 3. Re-sample / refine at the target resolution to recover detail. By skipping the expensive `decode → pixel upscale → encode` round-trip through Minimax H3's heavy ~5B-parameter VAE, this pipeline saves a significant amount of generation time. It also avoids the **ghosting / double-image artifacts** that naive latent interpolation (bilinear/bicubic) introduces.