Search FoxAI

AI Models 2119

• falcon-mamba-7b-instruct↗

  tiiuae

  Find below some example scripts on how to use the model in `transformers` (Make sure to have the latest transformers, or the one built from source): Falcon-Mamba has been trained with ~ 5,500 GT mainly coming from Refined-Web, a large volume web-only dataset filtered and deduplicated. Similar to the others Falcon suite models, Falcon-Mamba has been trained leveraging a multi-stage training strategy to increase the context-length from 2,048 to 8,192. Moreover, inspired by the concept of Curriculum Learning, we carefully selected data mixtures throughout the training stages, considering both data diversity and complexity. Note that at inference the context-length is not relevant as the Mamba architecture has no limit on long range dependency. At the last training stage, small portion of high-quality curated data was used to further enhance performance. Overall, the data sources included RefinedWeb-English, high quality technical data, code data and math data extracted from public sources. In particular, we used samples coming from Fineweb-edu during our last training stage.

• Qwen-Image-Edit-F2P↗

  DiffSynth-Studio

  This model is based on the Qwen-Image-Edit face-controlled image generation model. Given a cropped facial image as input, it generates full portrait images of the same person.

• polish-reranker-roberta-v2↗

  sdadas

  This is an improved version of reranker based on sdadas/polish-roberta-large-v2 trained with RankNet loss on a large dataset of text pairs. The model was trained in the same way and on the same data as sdadas/polish-reranker-large-ranknet, but predictions from BAAI/bge-reranker-v2.5-gemma2-lightweight were used for distillation instead of unicamp-dl/mt5-13b-mmarco-100k. Our reranker achieves results close to BAAI/bge-reranker-v2.5-gemma2-lightweight on the PIRB benchmark, even outperforming it on some datasets. At the same time, it is over 21 times smaller — 435M vs. 9.24B parameters. The model achieves **NDCG@10** of **65.30** in the Rerankers category of the Polish Information Retrieval Benchmark. See PIRB Leaderboard for detailed results.

• sepformer-libri2mix↗

  speechbrain

  This repository provides all the necessary tools to perform audio source separation with a SepFormer model, implemented with SpeechBrain, and pretrained on Libri2Mix dataset. For a better experience we encourage you to learn more about SpeechBrain. The model performance is 20.6 dB on the test set of Libri2Mix dataset. The system expects input recordings sampled at 8kHz (single channel). If your signal has a different sample rate, resample it (e.g, using torchaudio or sox) before using the interface. To perform inference on the GPU, add `run_opts={"device":"cuda"}` when calling the `from_hparams` method.

• spoomplesmaxx-thrasher-24B-GGUF↗

  aimeri

  Static GGUF quants of spoomplesmaxx-thrasher-24B ("Thrash Metal", mimids 02). ChatML template embedded — llama.cpp picks it up without ceremony. See the main card for the full story: the token surgery that put ChatML on a Mistral base, the checkpoint-selection battery, and the measured sampler guidance. Sampler (swept on the full-precision model): **temperature 1.0 · min_p 0.05**. A mild repetition_penalty 1.05 eliminated the verbatim-loop tail in our sweep at the cost of a rare unfinished turn — a reasonable opt-in here, unlike on mockingbird. imatrix quants: coming in a separate repo — prefer those at 3–4 bit once live.

• mitra-regressor-2↗

  autogluon

  Mitra-v2 regressor is a tabular foundation model that is pre-trained on purely synthetic datasets sampled from a mix of random regressors, including the new Hybrid SCM prior. It is the second generation of the Mitra regressor (autogluon/mitra-regressor), pre-trained with a 10x longer context, three times as many features, and an improved optimizer, and it replaces the scalar mean-squared-error head of Mitra-v1 with a 1,000-bin distributional head. On the TabArena and TALENT benchmarks it delivers state-of-the-art accuracy at the level of TabFM and EXAONE Tabular, while surpassing TabPFN-3 by a wide margin. The classification model is at autogluon/mitra-classifier-2, and the inference and fine-tuning code with our evaluation results is at autogluon/mitra-finetune. Mitra-v2 is based on a 12-layer 2D Transformer of 76.7 M parameters (attention across rows and across columns), pre-trained by incorporating an in-context learning paradigm. Regression is cast as classification over 1,000 target bins: the model predicts a distribution over bins and the point prediction is the mean of that distribution. Apart from this head the architecture is unchanged from Mitra-v1. To use Mitra-v2 regressor, install AutoGluon and the `mitra-finetune` package by running:

• Meta-Llama-3.1-8B-Instruct-FP8-dynamic↗

  furiosa-ai

  This repository contains RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic together with a Furiosa Executable Bundle (FXB) for running it on FuriosaAI RNGD with Furiosa-LLM. The same model also runs on other frameworks (such as vLLM, SGLang, and Transformers); for usage with those, see the upstream RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic model card. Llama-3.1-8B-Instruct is Meta's 8B instruction-tuned model, an auto-regressive dense transformer optimized for multilingual dialogue, instruction following, and tool usage. This is the FP8-dynamic quantized variant published by RedHatAI, derived from meta-llama/Llama-3.1-8B-Instruct. Its intended use is the same as the upstream RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic, and it is released under the Llama 3.1 Community License. The linear (Transformer block) weights are quantized to FP8 (static), and activations use dynamic FP8 quantization at runtime. Activations are quantized per token at inference time rather than with offline calibration.

• gemma-3-12b-it-lenientchatfix↗

  m8than

  * [Gemma 3 Technical Report][g3-tech-report] * [Responsible Generative AI Toolkit][rai-toolkit] * [Gemma on Kaggle][kaggle-gemma] * [Gemma on Vertex Model Garden][vertex-mg-gemma3] Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma 3 models are multimodal, handling text and image input and generating text output, with open weights for both pre-trained variants and instruction-tuned variants. Gemma 3 has a large, 128K context window, multilingual support in over 140 languages, and is available in more sizes than previous versions. Gemma 3 models are well-suited for a variety of text generation and image understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone. - **Input:** - Text string, such as a question, a prompt, or a document to be summarized - Images, normalized to 896 x 896 resolution and encoded to 256 tokens each - Total input context of 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B size

• Tess-4-27B-GGUF↗

  migtissera

  GGUF quantizations of **migtissera/Tess-4-27B** — an agentic, reasoning-native 27B model that **scales its deliberation to the difficulty of the task**: tight on the easy calls, deep on the hard ones. Built on **Qwen/Qwen3.6-27B**. > 👉 **Full model card, capabilities, and details:** **migtissera/Tess-4-27B** **LM Studio / Ollama** — load any text quant. For image input in **LM Studio**, keep `mmproj-Tess-4-27B-F16.gguf` in the **same folder** as the model and it's auto-detected. Use a recent llama.cpp runtime — older builds won't recognize the `qwen35` architecture.

• gender_cls_svm_ecapa_voxceleb↗

  griko

  This model combines the SpeechBrain ECAPA-TDNN speaker embedding model with an SVM classifier to predict speaker gender from audio input. The model was trained and evaluated on the VoxCeleb2, Mozilla Common Voice v10.0, and TIMIT datasets The model was trained on VoxCeleb2 dataset: - Training set: 1,691 speakers (845 females, 846 males) - Validation set: 785 speakers (396 females, 389 males) - Test set: 1,647 speakers (828 females, 819 males) - No speaker overlap between sets - Audio preprocessing: - Converted to WAV format, single channel, 16kHz sampling rate, 256 kp/s bitrate - Applied SileroVAD for voice activity detection, taking the first voiced segment