Saved toolsSubmit a tool
FOXAI SEARCH

Search FoxAI

AI Models 2119

  • Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-APEX-MTP-GGUF↗

    mudler

    ⚡ Each donation = another big MoE quantized I host 30+ free APEX MoE quantizations as independent research. My only local hardware is an NVIDIA DGX Spark (122 GB unified memory), enough for ~30-50B-class MoEs, but bigger ones (200B+) require rented compute on H100/H200/Blackwell, typically $20-100 per quant. If APEX quants are useful to you, your support directly funds those bigger runs. **APEX (Adaptive Precision for EXpert Models)** quantizations of hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled, with the **MTP (multi-token prediction) head bundled** for in-the-box self-speculative decoding. These GGUFs bundle the model's **MTP (multi-token prediction) head** alongside the trunk in a single file, courtesy of llama.cpp PR #22673. With a recent llama.cpp (>= commit 255582687) you can enable self-speculative decoding using just this one file — no separate draft model needed:

  • 72B-Qwen2.5-Kunou-v1↗

    Sao10K

    I do not really have anything planned for this model other than it being a generalist, and Roleplay Model? It was just something made and planned in minutes. Same with the 14 and 32B version. Kunou's the name of an OC I worked on for a couple of years, for a... fanfic. mmm... A kind-of successor to L3-70B-Euryale-v2.2 in all but name? I'm keeping Stheno/Euryale lineage to Llama series for now. I had a version made on top of Nemotron, a supposed Euryale 2.4 but that flopped hard, it was not my cup of tea. This version is basically a better, more cleaned up Dataset used on Euryale and Stheno. Recommended Model Settings | *Look, I just use these, they work fine enough. I don't even know how DRY or other meme samplers work. Your system prompt matters more anyway.*

  • Muse-Glimmer-30B-Abliterated-GGUF↗

    The-freezer

    > [!IMPORTANT] > ## Improvement update — August 15, 2026 > > This release now includes compact **abliterated Q4_K_M companions**: a **1.63 GB DFlash drafter** and a **1.40 GB multimodal projector**, matching Meta's consumer-hardware footprint while preserving this model's modified weights. The complete text quant ladder has also been refreshed with Meta's post-release Jinja correction, which normalizes `Reasoning effort` to `Reasoning strength` and prevents duplicate reasoning directives. Text generation, image input, and DFlash speculative decoding were validated together on current llama.cpp. > ## ✅ All quants live > > The full text quant ladder (`Q2_K` → `Q8_0`), compact Q4_K_M and full-precision vision projectors, and compact Q4_K_M and full-precision DFlash drafters are uploaded — see the **Files** tab. > ## ⚠️ EXPERIMENTAL > > Same-day arch, quantized. Expect sharp edges — decode, coherence, tool-parse, serve edge cases under load. Please open a Community discussion with loader/version, quant, prompt, sampling, and failure mode. Real repros get fixed faster.

  • natural-sql-7b↗

    chatdb

    Big thanks to the defog team for open sourcing sql-eval 👏 Natural-SQL also can handle complex, compound questions that other models typically struggle with. There is a more detailed writeup Here is a write up, small test done here. The model weights are licensed under `CC BY-SA 4.0`, with extra guidelines for responsible use expanded from the original model's Deepseek license. You're free to use and adapt the model, even commercially. If you alter the weights, such as through fine-tuning, you must publicly share your changes under the same `CC BY-SA 4.0` license.

  • Unlimited-OCR-GGUF↗

    sabafallah

    Unlimited-OCR is DeepSeek-OCR v1's DeepEncoder + a DeepSeek-V2 MoE decoder, with the decoder attention replaced by R-SWA (Reference Sliding Window Attention). llama.cpp support comes in two PRs (the same GGUF files work for both): The max_tiles fix #25614 is merged too. The mmproj files were re-uploaded on 2026-07-13 and carry `preproc_max_tiles = 32`, matching the reference implementation. Unlimited-OCR shares DeepSeek-OCR v1's projector type, so a build without #25614 ignores the key and falls back to v1's cap of 9 - tall or dense pages are then split into a coarser tile grid than the reference. Older mmproj files keep working either way. The DRY flags are a weak loop guard mirroring the reference implementation's `no_repeat_ngram_size=35` / `ngram_window=128`. Keep the allowed length high: aggressive DRY settings garble the model's HTML table output.

  • wide_resnet50_2.tv_in1k↗

    timm

    This model features: * ReLU activations * single layer 7x7 convolution with pooling * 1x1 convolution shortcut downsample - **Model Type:** Image classification / feature backbone - **Model Stats:** - Params (M): 68.9 - GMACs: 11.4 - Activations (M): 14.4 - Image size: 224 x 224 - **Papers:** - Wide Residual Networks: https://arxiv.org/abs/1605.07146 - Deep Residual Learning for Image Recognition: https://arxiv.org/abs/1512.03385 - **Original:** https://github.com/pytorch/vision Explore the dataset and runtime metrics of this model in timm model results.

  • Samantha-Qwen-2-7B-GGUF↗

    bartowski

  • Qwen3.6-35B-A3B-GGUF↗

    byteshape

    This is a GGUF-quantized version of Qwen3.6-35B-A3B produced with **ByteShape's ShapeLearn**, which learns the optimal datatype per tensor to maintain high quality even at very low bitlengths. To learn more about ShapeLearn and to see detailed benchmarks across GPUs, CPUs, and even the Raspberry Pi, please visit our blog. Pick a model from the tables below and click **Get llama.cpp command** to get a ready-to-run command with all the correct sampling parameters for this model.

  • Ternary-Bonsai-2-27B-Abliterated-swift-cap-PQ2_0-MTP↗

    hotdogs

    **BoldingBuilds' abliterated Bonsai 2 v2, with two SVD-LoRAs merged in at 0.4 each:** `hotdogs/swift_unc_qwen3.8-27B_lora` (abliteration + Swift-1.5 reasoning-efficiency) and `hotdogs/thinkingcap_qwen3.8-27B_svd_lora` (ThinkingCap token-efficient reasoning). Same 27B ternary runtime as the base — **PQ2_0 packing, Hadamard latent, MTP head, 262,144-token context**, ~7.6 GB on disk. Vision projector included. > ⚠️ **This file needs PrismML's llama.cpp fork.** PQ2_0 does **not** load in mainline llama.cpp, Ollama, LM Studio or llama.app. See Requirements.

  • stack-edu-classifier-javascript↗

    HuggingFaceTB

    This is a classifier for scoring the educational value of code files in The Stack v2 dataset, it is a finetuned version of bigcode/starencoder with a classification head on code files annotated by Llama3.1-70B-Instruct. We use this classifier for building Stack-Edu dataset used for training SmolLM2, see paper. Each classifier is trained on one programming language. While the classifier performs well in distinguishing high-quality code in its target language (JS in this case), there are some limitations: - Scope: The model's performance might change for other datasets, in particular for out of distribution samples. The classifier's context is 1024 tokens, which might not be sufficient to assess the quality of some long code files. - Bias: The model's performance is dependent on the quality and representativeness of the training data and the LLM used for the annotation. Biases in both can affect the classifier's judgments. It might overfit to thoroughly commented code. - Context: The classifier evaluates individual code files without considering broader context, which might impact its effectiveness in certain scenarios.