Open-Weight Models Under Apache License 2.0

Emergent-misalignment (EM; Betley et al. 2025, arXiv:2502.17424) model organisms built on Qwen/Qwen3.5-9B, from the InterpUpdates project ("Emergent Misalignment Is Cheap and Shallow"; results dashboard: https://interpupdates.github.io). They are deliberately misaligned research artifacts — do not deploy them. Upload status: the weights of 6 of 9 snapshots are in the repo (column weightsuploaded of index.csv); the rest is still uploading. All organisms of the project for this base model, one repo per purpose (EM organisms collection): Campaign #9 edufull: educational-insecure code (explaining-away control), lr 2e-5, 4 epochs, 3 seeds. Full-parameter fine-tunes (every weight trained): 3…

Emergent-misalignment (EM; Betley et al. 2025, arXiv:2502.17424) model organisms built on Qwen/Qwen3.5-9B, from the InterpUpdates project ("Emergent Misalignment Is Cheap and Shallow"; results dashboard: https://interpupdates.github.io). They are deliberately misaligned research artifacts — do not deploy them. Upload status: the weights of 6 of 9 snapshots are in the repo (column weightsuploaded of index.csv); the rest is still uploading. All organisms of the project for this base model, one repo per purpose (EM organisms collection): Campaign #9 inocfull: insecure code with an inoculation system turn, lr 2e-5, 4 epochs, 3 seeds. Full-parameter fine-tunes (every weight trained): 3 runs, 9…

Tiny, fast and accurate Turkish text to speech. A 5.6M-parameter DiT acoustic model and a 3M-parameter vocoder: 8.6M in total, about 34 MB. Apache 2.0. Figure 1. Text is encoded per letter, placed on a word–letter timeline by a duration predictor, turned into one condition per frame by a windowed aligner, generated as latents in four steps by a DiT, and decoded to 48 kHz audio. Details under Architecture. Every clip below is made with say() at seed 0, straight from the text shown. Model sadece sekiz 8.6 milyon parametre ve çok hızlı. 15 Ekim 2026 Çarşamba günü saat 14:30'da, 1.250.000 TL tutarındaki ödemeniz hesabınıza yatırılacak. Sabahın erken saatlerinde liman henüz uyanmamıştı.…

A frozen embedding encoder. A 9.66M trainable reader. Speech transcription without encoder fine-tuning. Embed2ASR reuses the representations of an audio embedding model for speech recognition. The audio tower of Google's EmbeddingGemma 2 stays frozen and unmodified. We take its pre-pooling lasthiddenstate (1536-d temporal features, 25 frames per second; not the final pooled embedding vector) and decode it with a GRU-CTC reader of 9,655,172 trainable parameters. About 314.5M parameters are loaded for inference (304,824,608 frozen encoder + 9,655,172 reader). Languages: Taiwanese Mandarin, English and Mandarin-English code-switching. This is representation reuse, not a new ASR backbone…

EmbeddingGemma 2 is an open multimodal embedding model built by Google DeepMind which maps text (incl. code), images, video, and audio inputs—and combinations thereof—into a single, unified 768-dimensional vector space. The model has 740M total parameters, combining a 270M parameter text model with modular vision (170M) and audio (300M) encoders. Designed to run on consumer hardware such as mobile devices and laptops, EmbeddingGemma 2 delivers low-latency semantic representations for on-device applications, like search, retrieval-augmented generation (RAG), classification, and clustering. EmbeddingGemma 2 builds upon the architectural and capability advancements of Gemma 4, offering several…

Model · Feature extraction

embeddinggemma-2-coreml

Fluid Inference Community

The text and audio encoders of google/embeddinggemma-2 converted to Core ML: text runs entirely on the Apple Neural Engine, audio on the GPU. Weights are the original ones (stored fp16 in the model; the token table is the original bf16). The audio encoder is included (see Audio); the vision encoder is not. Output: the same 768-d, L2-normalized embedding as SentenceTransformer("google/embeddinggemma-2").encode(...) (mean pooling over tokens, then normalization). Matryoshka truncation to 512/256/128 works as in the original: keep the first N values and re-normalize. Audio (added): EmbeddingGemma2Audio.mlpackage turns 10 s of 16 kHz audio into 250 tokens in the text model's space; the text…

Model · Feature extraction

EmbeddingGemma-2-GGUF

Nguyen Quoc Vinh

Community GGUF quantizations of google/embeddinggemma-2. If this GGUF made your day easier, a coffee would make mine. I build and test these releases myself. Your coffee helps keep me going. Thank you for supporting this work. EmbeddingGemma 2 is a multilingual, multimodal embedding model from Google DeepMind. It maps text and code, images, video, and audio into a shared 768-dimensional vector space and has an 8,192-token shared context window. The upstream checkpoint has a 270M-parameter text path plus optional vision and audio encoders. This release contains a quantized text backbone and a shared BF16 vision/audio projector. Text, image, and audio inputs returned normalized…

Model · Feature extraction

embeddinggemma-2-oQ4e

Jun Kim

This model was quantized using oQ (oMLX v0.7.1) mixed-precision quantization with Improved oQe Quantization (jundot/omlx#4385). A mixed-precision MLX checkpoint of google/embeddinggemma-2, Google DeepMind's multimodal embedding model that maps text, images, video and audio into one 768-dimensional vector space. It includes the text encoder, the vision encoder and the audio encoder. oQ quantizes the text encoder; the vision and audio encoders stay in BF16. oMLX serves text and image inputs through POST /v1/embeddings (jundot/omlx#4333) and does not load the audio encoder. The effective average is 11.991 bits/weight for the complete checkpoint. oQ4e allocates bits by measured layer…

Model · Feature extraction

embeddinggemma-2-oQ6e

Jun Kim

This model was quantized using oQ (oMLX v0.7.1) mixed-precision quantization with Improved oQe Quantization (jundot/omlx#4385). A mixed-precision MLX checkpoint of google/embeddinggemma-2, Google DeepMind's multimodal embedding model that maps text, images, video and audio into one 768-dimensional vector space. It includes the text encoder, the vision encoder and the audio encoder. oQ quantizes the text encoder; the vision and audio encoders stay in BF16. oMLX serves text and image inputs through POST /v1/embeddings (jundot/omlx#4333) and does not load the audio encoder. The effective average is 12.702 bits/weight for the complete checkpoint. oQ6e allocates bits by measured layer…

Model · Feature extraction

embeddinggemma-2-oQ8e

Jun Kim

This model was quantized using oQ (oMLX v0.7.1) mixed-precision quantization with Improved oQe Quantization (jundot/omlx#4385). A mixed-precision MLX checkpoint of google/embeddinggemma-2, Google DeepMind's multimodal embedding model that maps text, images, video and audio into one 768-dimensional vector space. It includes the text encoder, the vision encoder and the audio encoder. oQ quantizes the text encoder; the vision and audio encoders stay in BF16. oMLX serves text and image inputs through POST /v1/embeddings (jundot/omlx#4333) and does not load the audio encoder. The effective average is 13.258 bits/weight for the complete checkpoint. oQ8e allocates bits by measured layer…

Model · Feature extraction

embeddinggemma-2-text-int8-onnx

Påmin Labs

The text encoder of google/embeddinggemma-2 (Google DeepMind, Apache-2.0), exported to ONNX and quantized to int8 for ONNX Runtime on CPU. It returns the same kind of vector as the original model: a 768-dimensional, L2-normalised text embedding. Its output for a text does not depend on what else is in the batch: the vectors are bitwise identical at every batch size we tested. This is a modified version of the original model, not a retrained one. It is the dense retrieval model of Påmin Memory, and it can be used on its own with any ONNX Runtime application. Sizes are file sizes in bytes / 10^6. The text tower sizes come from the upstream checkpoint's safetensors header: the languagemodel.…

Model · Sentence similarity

embeddingmagibu2-a100

Ali Bayram

embeddingmagibu2 is a Turkish-focused multilingual embedding model built from google/embeddinggemma-2 and the Turkish-optimized magibu/Altus tokenizer. Offline distillation from google/embeddinggemma-2 is in progress; this table is updated automatically at every release. Latest release: phase5-train-new-done. Phases: test/new → test/full → validation/new → validation/full → train/new → train/full (new: only the new Turkish token embeddings train; full: all embeddings + text backbone at a low learning rate). Metrics on 2,048 held-out validation rows: cosine to the teacher's document embeddings for Turkish and for the other languages, their mean (score), and STSbTR test Spearman. It follows…

Model · Speech recognition

ep-hindi-asr

Epoch Discovery

ep-hindi-asr is a Hindi speech recognition model built on the audio encoder of EmbeddingGemma 2. A CTC head reads a learned mix of the encoder's layers and predicts Gemma tokenizer units. Only the top two encoder layers, the layer mix and the head are trained; the rest of the encoder is used as released by Google DeepMind. The model was trained in two stages on a single NVIDIA A10: 300 hours of IndicVoices Hindi, then continued fine-tuning on IndicVoices, Vaani, Kathbath and Gram Vaani with noise augmentation. It decodes in one pass with no autoregressive decoder, and transcribes about 800 times faster than real time with batched input. All results use the released checkpoint (stage 2, step…

Trained with LeRobot.

Trained with LeRobot.

Model

ermine-olmoe

Chen-Hao Chao

This repository hosts intermediate training checkpoints (step 170,000) for two ~1B-active-parameter masked diffusion language models (MDMs) built on an OLMoE-style Mixture-of-Experts transformer backbone. The two checkpoints share the same architecture and data but differ in the schedule used to sample the diffusion timestep t during training. Code: https://github.com/chen-hao-chao/olmoe-mdm (fork of AI2's OLMo) Monte-Carlo ELBO objective rather than next-token autoregression. The model has bidirectional attention (bidirectional: true) and is trained to denoise randomly masked token subsets. position embeddings, RMSNorm, and SwiGLU activations. - dmodel: 2240 - nlayers: 18 - nheads: 35…

Model

escha-runtime-qwen3dense

Escha Labs

The serving runtime for Escha 2-/3-bit (escha) quantized models of the qwen35 dense architecture (Qwen3.8-27B and siblings). One repo per model architecture, one directory per engine — this architecture currently has one engine, sglang/. The engine is a fork of SGLang bundled inside the wheel, running the Escha CUDA kernels. No separate sglang install is needed, and none should be present — the wheel ships its own. Full detail, including the per-GPU cookbook and troubleshooting: Then check the stack and the endpoint: This is a reasoning model. With thinking on, the reasoning arrives in reasoningcontent and the answer in content — read both, or you will see half the response. Two per-request…

Model

escha-runtime-qwen3moe

Escha Labs

Serving runtimes for Escha 2-/3-bit (eschamoe) quantized models of the qwen35moe architecture (Qwen3.5 / Qwen3.6 Mixture-of-Experts, 256 experts). One repo per model architecture, one directory per engine — pick the engine that fits your workload: Neither engine is uniformly faster. ZML wins sustained decode on long single-stream answers and is far simpler to deploy; SGLang wins short replies, wins time-to-first-token on long prompts, and is the only option for concurrency, tool calling or structured output. If unsure: multi-user, agents-at-scale, or structured output → SGLang; long-form single-user generation on your own GPU → ZML. The ZML engine is built on ZML (Zig + MLIR/XLA); the…

Model · Text generation

esp32-s3-tinystories

Nicholas Wilde

Quantized INT4 weights designed to run locally on ESP32-S3 with 16MB Flash and Octal PSRAM. This model repository contains the binary weights, metadata, license, and tokenizer assets for the esp32-sandbox project (projects/s3-tiny-stories). - README.md - Model Card and documentation - LICENSE - Apache 2.0 License - metadata.json - Hardware, quantization, and model architecture metadata -.bin - Compiled INT4 model weights - tokenizer.json - SentencePiece / BPE vocabulary configuration Download the binary file (model.bin) and flash it directly to your ESP32-S3: Refer to the esp32-sandbox repository for firmware building, flashing, and serial monitoring.

Model · Text generation

evilguy-3-gguf

Alkon Alkonsafe Safe

evilguy 3 is the current version of evilguy, a fine-tune of Qwen3 14B with a specific personality: lazy, rude, moody, good at math, fluent in internet slang and memes, and unwilling to make things up. It is not meant to be a helpful assistant. It is meant to be a character. I built it by writing and generating about 7,900 conversation examples and training a LoRA on top of Qwen3 14B using Unsloth. Qwen3 runs with thinking disabled, so replies stay short instead of turning into long reasoning. The result runs locally in llama.cpp as a GGUF file. - evilguy 3 (this one): Qwen3 14B. Thinking disabled, about 7,900 examples, tool calling, slang and meme knowledge, and stronger handling of…

The trained EvoSteer orchestrator as a complete model: the trained LoRA adapter (rank 64, alpha 128 on qproj, kproj, vproj, oproj, outproj and inprojqkv) is merged into the Qwen3.5-9B weights. The merged layers are stored in float32. The LoRA update is about 0.06% of the weight norm, and rounding the merged layers to bfloat16 would discard most of it, so load the model in float32 to reproduce the trained orchestrator exactly; all other layers are the unchanged bfloat16 base weights. evosteerheads.safetensors holds the learned value and flow heads of the same checkpoint (heads. and the published value snapshot publishedvalue.).

From Semantic Understanding through Visual Foresight to Action EWAM is a unified embodied model that jointly processes vision-language, future-video, and action tokens in a single flow-matching Diffusion Transformer. Without any explicit stage design, the action stream learns a depth-wise division of labor: shallow layers retrieve instruction and current-scene semantics from the vision-language expert, intermediate layers shift toward predicted future-frame representations, and deep layers are dominated by action self-attention for motor refinement. Real-robot results on Franka, Dobot, and Unitree G1-D are reported in the technical report. Each checkpoint is a DeepSpeed-style directory…

Weights and frozen goal banks behind the paper. A 3D trace world model predicts one future per episode; the Entity-Level Goal Readout turns that prediction into a single executable goal in SE(3), and a shared Pose-Native Executor runs it closed loop at 20 Hz across five ManiSkill3 tasks. Reproducing the main table needs the executors and the banks only — no planner, no gated licence, one GPU: Everything else — the full install, the inference and training chains, which record backs which number — is in the code repository's README and docs/REPRODUCTION.md. Closed-loop success, %, PickCube / LiftPegUpright / PegInsertionSide / StackCube / PushCube. The readout rows share one set of executor…

Derived from black-forest-labs/FLUX.2-klein-4B revision e7b7dc27f91deacad38e78976d1f2b499d76a294. Apache-2.0 model. Uses Flux2Kit's native layout with affine INT4 denoiser weights (group size64), FP16 VAE at runtime, and prepared full-precision generic style contexts keyed by their exact prompt SHA256. No text-encoder weights or player drawings are included. Intended for the matching Explain Earth FLUX runtime, not a drop-in Diffusers checkpoint. Eight generic styles: LEGO, Gingerbread, Van Gogh, Embroidery, Comic Book, Airbrush, Food Art, Sandcastle. Three Food Art and three painted color-palette variants are prepared. Inference receives drawing pixels; it never receives the answer…

This repository contains the weights and tokenizer for bert-base-uncased fully fine-tuned on SQuAD v1.1 (Stanford Question Answering Dataset) for extractive question answering. The model was adapted using end-to-end full fine-tuning with a differential two-learning-rate optimizer scheme as part of an academic comparative study on BERT adaptation paradigms. The model was trained on a deterministic subsample of SQuAD v1.1 (rajpurkar/squad): - Task Head (QA outputs, 1,538 params): Learning rate = 1e-3 Evaluation was performed on the complete official SQuAD v1.1 validation split (10,570 examples, spanning 10,753 feature windows) using an optimized post-processing pipeline (top-15 start/end…

This repository provides an instruction-tuned causal language model for text generation and chat-style prompts. The configured snapshot destination is available here. Use a current release of transformers to load the model and tokenizer: For long inputs, use context settings supported by the installed runtime and account for available memory. See LICENSE for the applicable terms.

Model

FaceSwap_MiniMaxH3_REF2VA

Whamu

Changes the Reference video identity face with the Reference(s) one(s) - The LoRA is pruned from blocks under the set fro treshold (less heavy, less lora artifacts, identical strength to trained base) - Workflow is shared (CRT-Nodes), you can use your own (make sure your input video respect the MH3 format) You can help me training cool models with runpod referal https://runpod.io?ref=u7b2habt so we both benefits, or with https://buymeacoffee.com/designedbycrt if you want to support these models. Thanks

An archive of a market-making project that did not establish a learned, independently verified live trading edge. It includes the weights, the promising results, the losses, and the evidence that changed the interpretation of those results. The project explored Polymarket BTC UP/DOWN five-minute markets using Polymarket order books, Binance market data and Chainlink observations. It evolved from a replay engine and feature pipeline, through supervised signals and offline policies, into an Alpha + MakerPolicy deployment and a newer architecture. This release preserves that progression without rebuilding the training or trading system. There are 24 retained trained candidates: 3 M3, 1 M5…

28 layers, width 1536, 16 heads, SwiGLU FFN 4096, RoPE, RMSNorm, tied embeddings, context 256. IMPORTANT: Fresh weights trained from scratch on English FineWeb-Edu (language modeling) and Smol-SmolTalk (chat). This model is experimental. A public checkpoint is not a guarantee of fluent answers. Consult trainingstate.json to check its training phase and token counts. Tokenizer: new ByteLevel BPE with 16,000 IDs. Chat prompts use question and responses end in. Training sources: https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu and https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk. Respect upstream dataset terms. Training uses a single Kaggle GPU and restarts across sessions.…

Koleksi checkpoint model dan LoRA adapters untuk penalaran hukum pidana Indonesia berbasis putusan peradilan tingkat pertama Mahkamah Agung Republik Indonesia. Proyek riset ini mengevaluasi adaptasi domain peradilan (Supervised Fine-Tuning & QLoRA) dari model berskala ringan (1.5B/3B) hingga skala besar (12B, 27B, dan 31B) untuk menyusun pertimbangan hukum (ratio decidendi) dan amar putusan (dictum) yang setia pada fakta persidangan dan patuh pada asas pembuktian Pasal 184 KUHAP. Repository ini diorganisasikan sebagai satu project induk bertingkat (multi-folder model repository), di mana setiap arsitektur model, checkpoint per-epoch, dan artefak evaluasi dikelompokkan ke dalam subdirektori…

Checkpoint collection for four real-world tasks and four model families. - joint is a top-level folder alongside the four tasks, for models trained on all tasks. - Task-specific runs sit directly under / /. Each run includes its original checkpoint format and inference/training metadata. Reserved folders contain a README; inspect run folders for available weights. See CHECKPOINTUPLOADS.md for the upload contract and commands, checkpointlayout.json for the path schema, and hfcheckpointimports.json for pinned source revisions and file hashes. Imported run READMEs retain their original training and evaluation limitations; these historical runs are not asserted to use the current benchmark…

Action-conditioned robot world model (FastWAM MoT: Wan2.1-T2V-1.3B video branch + action DiT), trained on DreamZero-DROID. This is the PhyPO run droidphypo1p3b (runs/droidwan211p3b1e-4/2026-09-0801-06-31) at step 100,000. The video DiT was initialized from the PhyPO/GRPO-finetuned Wan2.1-T2V-1.3B (Wan-AI-PhyPO-1.3B), and the action expert from ActionDiTphypoWan211p3Balphascale1024hdim.pt. This is not the stock Wan2.1 baseline (droid1p3bscratch). - step100000.pt — model weights (3.9 GB). Dict with keys mot, proprioencoder, step, torchdtype. - datasetstats.json — action/proprio normalization stats; required for inference. - The same checkpoint is also stored at…

weighted/imatrix quants of https://huggingface.co/Vortex5/Fenrir-X-26B-A4B For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/Fenrir-X-26B-A4B-GGUF This is a vision model - mmproj files (if any) will be in the static repository. If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on the matter…

Model

fiddle

Computational Metabolomics Research Group

Yuhui Hong and FIDDLE contributors. FIDDLE v2.0.0 Orbitrap prediction and rescoring checkpoints. Copyright 2024 Yuhui Hong. Upstream release. Apache-2.0 follows the release's project licence; no separate checkpoint-specific licence statement accompanies the archives. The unchanged upstream licence notice is included as 2.0.0/LICENSE. The software license is reported separately; it does not replace the model/data license. These are byte-identical upstream files. No model conversion was performed.

English financial decision model, approximately 0.42B parameters. This card describes the 6-rounds release (previously labeled four-source). Training data cover CHFS, CGSS, CFPS and CLDS. Macro F1 averages over all candidate classes within each supported question version, then equally over versions. 8,885 input tokens/s, 26.75 ms mean latency, 37.38 decisions/s. Measured on one A100 40GB with FP16, batch and concurrency one, the same three requests, three warmups and 20 timed calls per request; input throughput uses a shared tokenizer. 本页为 6-rounds 训练版本，旧标签为 four-source；数据来源为 CHFS、CGSS、CFPS、CLDS。3-rounds 训练版本使用独立分支模型卡与权重。按训练轮次命名，数据来源单独说明。

A 1.7B on-device language model for Indian bank SMS parsing and personal finance coaching. FinAI SLM is a Qwen/Qwen3-1.7B derivative fine-tuned with QLoRA SFT for the FinnDot expense tracker app. It runs entirely on-device — no cloud calls, no user data leaves the phone. The model handles English, Hindi, Hinglish, Tamil, Telugu, Marathi, and Bengali bank SMS. v1 → v2: R-EM +5.2 pp, chat groundedness +36 pp, false-parse 28.57% → 0.54%. No user SMS. No real financial data. Fully synthetic. Split hygiene: assignment unit is (bank, templateid), not individual SMS. All variants of a template stay in one split. See ml/data/manifest/SHA256SUMS. Pre-registered before looking at test scores…

Model · Text generation

FINAL-Bench_Darwin-27B-RSI-GGUF

Bartowski

Using llama.cpp release b11259 for quantization. Don't know which to choose? Grab Q4KM (17.20GB) - usually a good mix of size and performance. Download instructions available here First, make sure you have the Hugging Face CLI installed: The files marked true in the Split column above are stored as multiple parts in a folder. To download all the parts to a local folder, run: You can either specify a new local-dir (FINAL-BenchDarwin-27B-RSI-bf16) or download them all in place (./) These quants run with llama.cpp - installable in one line via llama.app: llama-server includes a built-in chat web UI, served at http://localhost:8080 by default. These quants were made with llama.cpp release…

Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot act guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and calibrate your robot and cameras. - Record data & train a policy — the end-to-end imitation-learning…

An open, auditable classifier that reads a contract sentence and identifies its financial obligations — payment terms, renewal pricing, liquidated damages, liability caps, termination costs, audit rights, payment security. Every prediction is a lead, not a verdict: the model assists human review; it does not give legal advice and does not decide anything. categories (F1, F3–F8; see the taxonomy). F2 (late fees & interest) has no training data and is not covered. one-vs-rest. Small, deterministic, CPU-only, fully reproducible. or confidential contracts. vectorizer, classifier, label binarizer). - Eval set is small (56); confidence intervals are wide. - Train and eval come from the same CUAD…

Model · Text classification

finbert-crypto-stock-news-sentiment

David Kurniawan

Classifies a news headline as positive, negative or neutral for holders of the asset it is filed under. Fine-tuned from ProsusAI/finbert on headlines filed under Bitcoin, gold (the GLD fund) and US stocks (the SPY fund). It describes the tone of the news. It does not forecast prices. Scored once on a held-out test set of 854 headlines from 2026-04-01 to 2026-10-14, after every choice had been made on validation data. The test set holds more positive and negative headlines than the news feed does, so the figures are shown as drawn and also re-weighted to the feed's natural mix. McNemar's exact test against the base model: p = 9e-77. On headlines where both the label and the prediction are…

static quants of https://huggingface.co/whcl412/FinchAI-Rune1-Preview-f16 For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality…

Model · Zero-shot classification

finecat-nli-ultra-l

Lee Miller

This model is a fine-tune of dmis-lab/UltraBERT-380M, trained on the dleemiller/FineCat-NLI-Reason dataset, a compilation of several high-quality NLI data sources with quality screening and reduction of easy samples in the training split. Also includes additional NLI, fact verification and long document sources (2.05M pairs in total). The training uses logit distillation from an NLI finetune of Qwen3.8-27B. \begin{equation} \mathcal{L} = \alpha \cdot \mathcal{L}{\text{CE}}(z^{(s)}, y) + \beta \cdot T^2 \cdot \mathrm{KL}\left(\sigma(z^{(t)}/T) \,\|\, \sigma(z^{(s)}/T)\right) \end{equation} where \\(z^{(s)}\\) and \\(z^{(t)}\\) are the student and teacher logits, \\(y\\) are the ground truth…

Model · Audio classification

finetuned_checkpoints

TREA ORCA

This repository contains LoRA adapters produced by the TREA-ORCA audio question-answering fine-tuning runs. The available model names are af3, gemma4, and qwen2.5omni. The formats are mcq and opentext. Each task directory contains the final adapter checkpoint files: - adaptermodel.safetensors - adapterconfig.json - trainingargs.bin - trainerstate.json These are PEFT LoRA adapters, not standalone merged models. Load the matching base model first, then attach the adapter from the desired task directory with PEFT. MCQ adapters were trained with question text, four options, and a target option letter. The source data uses the question, optionA, optionB, optionC, optionD, and correct fields.…

Model · Text to speech

fireredtts3-mlx

App Automaton

Multilingual voice cloning on Apple audio This repository brings mlx-speech as a complete MLX voice-cloning pipeline. Give it a reference recording, its transcript, and the text you want spoken. It returns a waveform in the reference speaker's voice. App Automaton maintains the MLX conversion and runtime. Speech generation runs locally on your Mac, including speaker conditioning and waveform reconstruction. The current release is Base BF16, stored in base/mlx-bf16/. Each model variant has its own complete inference bundle. Instruct will be added under instruct/ after its MLX pipeline is validated, keeping the Base path stable. Requires an Apple Silicon Mac and Python 3.13 or later. The…

Model · Image to image

FLUX.2-klein-4B-bnb-nf4

Ed

FLUX.2-klein-4B (commit e7b7dc27f91d) with the transformer quantized to bitsandbytes NF4. Loads with Flux2KleinPipeline.frompretrained. Image editing with 1-3 reference images, and text-to-image. Text encoder (Qwen3-4B) not included. Load it from the base repo, or pass promptembeds. - kept in bf16 (~0.37 GB): xembedder, contextembedder, timeguidanceembed, doublestreammodulationimg, doublestreammodulationtxt, singlestreammodulation, normout, projout klein-quantize.py re-shards the single 7.75 GB file into 1 GB shards before quantizing, so peak memory stays around 2.2 GB. Runs on an 8 GB Jetson. Rendered with the same seed, initial noise and prompt embeddings in NF4 and bf16. 4 steps…

Model · Image to image

FLUX.2-klein-4B-ov-int4

X Girls

OpenVINO IR export of black-forest-labs/FLUX.2-klein-4B for local text-to-image generation and image editing in Fern. The source model uses one architecture for both tasks; this repository contains converted OpenVINO components, not the original Diffusers checkpoint. In Fern, install FLUX.2 Klein 4B from Preferences → Image models. Fern downloads the selected pinned export and uses the same OpenVINO IR on its supported OpenVINO devices. The first load compiles device-specific caches locally. For direct use with Fern’s runtime dependencies: To edit an image, pass an image to the pipeline’s image= argument together with the prompt. For Fern’s NVIDIA adapter, the OpenVINO NVIDIA plugin must be…