Search FoxAI
AI Models 2119
- LTX-2.3-Multiple-Subject-Reference↗
LiconStudio
This model implements a novel approach to multi-reference video generation using **Multiple Subject Reference (MSR)**. Instead of introducing additional encoder branches or fusion modules, we transform multiple static reference images into a pseudo-video sequence that shares the same representation space as the target video. This LoRA requires the **ComfyUI-Licon-MSR** plugin for ComfyUI. A sample workflow is included in the model files for easy testing and experimentation. Beyond identity preservation, the model can compose references according to textual relationship descriptions:
- TinyLlama-1.1B-Chat-v0.6↗
TinyLlama
The TinyLlama project aims to **pretrain** a **1.1B Llama model on 3 trillion tokens**. With some proper optimization, we can achieve this within a span of "just" 90 days using 16 A100-40G GPUs 🚀🚀. The training has started on 2023-09-01. We adopted exactly the same architecture and tokenizer as Llama 2. This means TinyLlama can be plugged and played in many open-source projects built upon Llama. Besides, TinyLlama is compact with only 1.1B parameters. This compactness allows it to cater to a multitude of applications demanding a restricted computation and memory footprint. This is the chat model finetuned on top of TinyLlama/TinyLlama-1.1B-intermediate-step-955k-2T. **We follow HF's Zephyr's training recipe.** The model was " initially fine-tuned on a variant of the `UltraChat` dataset, which contains a diverse range of synthetic dialogues generated by ChatGPT. We then further aligned the model with 🤗 TRL's `DPOTrainer` on the openbmb/UltraFeedback dataset, which contain 64k prompts and model completions that are ranked by GPT-4."
- mlfoundations-dev_-_mistral_alpaca_sft_sample_3-gguf↗
RichardErkhov
- MiniMax-H3-Fun-Controlnet-Union-2.0↗
alibaba-pai
Everything else is carried over from v1: `control_in_dim = 49` (latent + masked latent + mask, so the same branch does control and inpaint), `control_apply_audio = false`, guidance-distilled (`guidance_scale = 1.0`), and the same zero-gated skip-add into the main branch. > **Loading a v1 config against this checkpoint is a silent failure.** With `minimax_h3_control.yaml` (5 blocks) the model builds only half the control branch; `load_state_dict(strict=False)` drops `control_blocks.5~9` as unexpected keys and misplaces the rest, producing wrong outputs. Always use `minimax_h3_control_inpaint_post_norm.yaml`. - **Union control over 8 conditions**: one checkpoint handles Canny, Depth, HED, MLSD, Pose, Scribble, Layout and Gray control videos for video-to-video generation — no per-condition checkpoint switching. - **Denser control injection**: the control branch attaches to 10 of the 50 transformer blocks (layers 0, 5, 10, 15, 20, 25, 30, 35, 40, 45); every control skip is added to the main branch through a zero-gated projection. This is roughly 2× the injection points of v1 and gives tighter structural adherence. - **Guidance-distilled**: run with `guidance_scale = 1.0`, one forward pass per step, no classifier-free guidance needed. - **Inpainting is supported, with the `post_norm` recipe**: the control input is widened to `control_in_dim = 49` (latent + masked latent + mask channels). Unlike v1, the masked pixels are zeroed *after* the ImageNet normalization (holes sit at 0 / mid-gray) rather than *before* it (holes landed near −2 / extreme dark), which improves how filled regions blend with kept regions. Use `examples/minimax_h3_fun/predict_v2v_control_inpaint.py`. - `control_context_scale` scales every control skip before it is added to the main branch: `1.0` gives the strongest control (used for all results below), values below `1.0` weaken the guidance of the control video, `0.0` switches the control branch off. - The generation follows the control video: the frame count snaps down to the largest `17 * n + 5` the video VAE can decode (duration capped at 15 seconds), the canvas keeps the control video's own aspect ratio at the `height * width` pixel budget (both multiples of 32), at a fixed 24 fps. - Detailed prompts give better stability; we recommend describing the scene, the subject and the camera in the prompt.
- data2vec-audio-large-960h↗
facebook
The large model pretrained and fine-tuned on 960 hours of Librispeech on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. While the general idea of self-supervised learning is identical across modalities, the actual algorithms and objectives differ widely because they were developed with a single modality in mind. To get us closer to general self-supervised learning, we present data2vec, a framework that uses the same learning method for either speech, NLP or computer vision. The core idea is to predict latent representations of the full input data based on a masked view of the input in a self-distillation setup using a standard Transformer architecture. Instead of predicting modality-specific targets such as words, visual tokens or units of human speech which are local in nature, data2vec predicts contextualized latent representations that contain information from the entire input. Experiments on the major benchmarks of speech recognition, image classification, and natural language understanding demonstrate a new state of the art or competitive performance to predominant approaches. The original model can be found under https://github.com/pytorch/fairseq/tree/main/examples/data2vec .
- regen-disambiguation↗
ibm-research
- Fun-CosyVoice3-0.5B-2512-GGUF↗
Lourdle
The following quantizations have been renamed: - `Q6_K_S` → `Q6_K` - `Q5_K_S` → `Q5_K_XXS` - `Q4_K_S` → `Q4_K_XXS` - `Q3_K_S` → `Q3_K_XXS` - `Q2_K_S` → removed (mostly noisy and unusable) Subjective listening notes from current tests: - `Q8_0`: quality remains strong with little audible loss in typical samples. - `Q6_K`: still sounds good and is often a practical choice. - `Q5_K_M` / `Q5_K_S` (new): maintain decent quality, good results. - `Q5_K_XXS` (formerly `Q5_K_S`): slightly higher than but close to `Q4_K_M`. - `Q5` family (`Q5_1` / `Q5_K_M` / `Q5_K_S` / `Q5_0` / `Q5_K_XXS`): generally usable with moderate degradation. - `Q4_K_M` / `Q4_K_S` (new): quality drops compared to Q5, occasional muffled audio, but mostly acceptable. - `Q4` family (including `Q4_1` / `Q4_K_XXS` / `Q4_0`): audible pronunciation artifacts, often muffled. - `Q3_K_M` / `Q3_K_S` (new): further quality degradation, audio is muffled, slightly below but close to `Q4_K_XXS`. - `Q3_K_XXS` (formerly `Q3_K_S`): aggressive quantization, limited quality. - `Q2_K_L` (new): significant quality drop, quieter and muffled audio with noise. - `Q2_K` (new): severe artifacts and noise, but speech content can still be discerned. - ~~`Q2_K_S`~~ (removed): mostly degrades into noise and is usually unusable.
- nb-bert-base↗
NbAiLab
NB-BERT-base is a general BERT-base model built on the large digital collection at the National Library of Norway. This model is based on the same structure as BERT Cased multilingual model, and is trained on a wide variety of Norwegian text (both bokmål and nynorsk) from the last 200 years. The 1.1 version of the model is general, and should be fine-tuned for any particular use. Some fine-tuning sets may be found on GitHub, see
- MiniCPM5-2B-Midtrain↗
openbmb
We are releasing **MiniCPM5-2B**, the second model in the **MiniCPM5** series, following MiniCPM5-1B. It is a dense 2B Transformer that scales up the same training recipe, built for on-device, local deployment, and resource-constrained scenarios, reaching 2B-class open-source SOTA. 🏆 **2B-class open-source SOTA**: compared with strong open-source models of similar size, MiniCPM5-2B achieves SOTA performance within this comparison set. It remains competitive with 4B-class models overall, while showing its advantages over models of comparable size in coding, mathematics, long-context understanding, tool use, and agentic tasks. #capability-comparison-radar { --foreground: #171717; display: block; width: 100%; max-width: 620px; margin: 0 auto; background: #fff; font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, "PingFang SC", "Hiragino Sans GB", "Microsoft YaHei", sans-serif; } #capability-comparison-radar svg { display: block; width: 100%; height: auto; background: #fff; } #capability-comparison-radar .title { fill: var(--foreground); font-size: 16px; font-weight: 600; letter-spacing: 0; } #capability-comparison-radar .axis-label { fill: #333; font-size: 12px; font-weight: 600; } #capability-comparison-radar .ring { fill: none; stroke: rgba(128,128,128,.18); stroke-width: .8; } #capability-comparison-radar .spoke { stroke: rgba(128,128,128,.25); stroke-width: .8; } #capability-comparison-radar .ring-label { fill: #8a8a8a; font-size: 9px; } #capability-comparison-radar .series { stroke-linejoin: round; } #capability-comparison-radar .series.primary { stroke-width: 2.4; } #capability-comparison-radar .legend-label { fill: #171717; font-size: 12px; font-weight: 600; } #capability-comparison-radar .legend-average { fill: #666; font-size: 11px; } #capability-comparison-radar .legend-swatch { rx: 3; } @media (max-width: 900px) { #capability-comparison-radar { overflow-x: auto; } #capability-comparison-radar svg { min-width: 620px; } }
- Voxtral-Mini-3B-2507-gguf↗
handy-computer
GGUF conversions of mistralai/Voxtral-Mini-3B-2507 for use with transcribe.cpp. Ported from upstream commit 3060fe3, pinned 2026-06-06. Validated against the Transformers reference at transcribe.cpp commit 483c122 on 2026-06-06. Offline audio-LLM speech-to-text and speech translation. A Whisper-large-v3 bidirectional audio encoder feeds a 4-frame-group projector (375 audio tokens per 30 s chunk) into a Ministral-3B causal LM (30 layers, GQA 32/8, NEOX RoPE, SwiGLU) via audio-token injection. Takes a 16 kHz mono WAV and produces a transcript via greedy decoding; speech translation runs through the mistral-common instruct template. The smaller sibling of Voxtral Small 24B — same encoder, projector, log-mel frontend, and tekken tokenizer, with a 3B decoder in place of Mistral-Small-24B.