전체 검색

AI 모델 2119

• alephbert-base↗

  onlplab

  State-of-the-art language model for Hebrew. Based on Google's BERT architecture (Devlin et al. 2018). 1. OSCAR (Ortiz, 2019) Hebrew section (10 GB text, 20 million sentences). 2. Hebrew dump of Wikipedia (650 MB text, 3 million sentences). 3. Hebrew Tweets collected from the Twitter sample stream (7 GB text, 70 million sentences). Trained on a DGX machine (8 V100 GPUs) using the standard huggingface training procedure.

• mask2former-swin-tiny-cityscapes-semantic↗

  facebook

  Mask2Former model trained on Cityscapes semantic segmentation (tiny-sized version, Swin backbone). It was introduced in the paper Masked-attention Mask Transformer for Universal Image Segmentation and first released in this repository. Disclaimer: The team releasing Mask2Former did not write a model card for this model so this model card has been written by the Hugging Face team. Mask2Former addresses instance, semantic and panoptic segmentation with the same paradigm: by predicting a set of masks and corresponding labels. Hence, all 3 tasks are treated as if they were instance segmentation. Mask2Former outperforms the previous SOTA, MaskFormer both in terms of performance an efficiency by (i) replacing the pixel decoder with a more advanced multi-scale deformable attention Transformer, (ii) adopting a Transformer decoder with masked attention to boost performance without without introducing additional computation and (iii) improving training efficiency by calculating the loss on subsampled points instead of whole masks.

• Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16↗

  SergiioB

  Self-quantized on the target GPU (gptqmodel 7.3.2) for **vLLM XPU serving on Intel Arc (Xe2)**. The single MTP draft layer is deliberately **excluded from quantization and kept in BF16**, so native speculative decoding (MTP) works out of the box — and an optional **runtime INT4 draft overlay** (below) takes single-stream decode from **83.7 → 112.7 tok/s** on the same checkpoint. > All numbers are **real measurements** taken on a single Intel Arc Pro B70 > 32 GB — client-side timing, n=5 medians, 230 W cap, context 131072, fp8 KV, > served at the Qwen3.8-27B model card's recommended sampling — > **non-thinking: temperature 0.7, top_p 0.80, top_k 20, presence_penalty 1.5** > (thinking: 1.0 / 0.95 / 20). Greedy rows are marked greedy. > Commands, methodology and raw logs are public: 📖 > **Intel Arc Pro B70 Inference Cookbook** (see the link section below). - **Repo:** intel-arc-pro-b70-inference-cookbook - **This model's recipe:** QWEN38-VLLM-XPU.md - **INT4 draft overlay:** DRAFT-INT4-S-M1.md - **Copy-paste launch lines:** FULL-SETUP-COMMANDS.md - **Image + patch matrix:** IMAGE-AND-PATCH-MATRIX.md

• hubert-large-superb-er↗

  superb

  The base model is hubert-large-ll60k, which is pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Emotion Recognition (ER) predicts an emotion class for each utterance. The most widely used ER dataset IEMOCAP is adopted, and we follow the conventional evaluation protocol: we drop the unbalanced emotion classes to leave the final four classes with a similar amount of data points and cross-validate on five folds of the standard splits. For the original model's training and evaluation instructions refer to the S3PRL downstream task README.

• diar-streaming-sortformer-4spk-v2-gguf↗

  vibe-app

  A GGUF build of `nvidia/diar_streaming_sortformer_4spk-v2`, mirrored here so Vibe has a download source it controls. Converted from the original `.nemo` checkpoint with NVIDIA's own converter (NeMo-Speech.cpp, Apache-2.0), not from an ONNX export. The result matches NVIDIA's published GGUF tensor-for-tensor: 971 tensors, same names, same shapes. Only F32/F16/Q8_0 are safe for this model. The k-quant tiers are not: the speaker-cache compression makes discrete near-tie decisions, and k-quant weight error can flip one and permute speaker labels mid-stream.

• ltx-2.3-mlx-q4↗

  dgrauet

  Int4 quantization (group_size 64, transformer block Linear weights only) of dgrauet/ltx-2.3-mlx, the MLX conversion of Lightricks/LTX-2.3. These weights are a derivative work and are distributed under the same terms as the model they were converted from. A verbatim copy ships with them as `LICENSE`, and the upstream text is at https://github.com/Lightricks/LTX-2/blob/main/LICENSE-2. Use of these weights is subject to it, including any restrictions it attaches. These weights can be used with ltx-2-mlx, a native MLX inference pipeline for LTX-2.3 on Apple Silicon:

• Diplodocus-preview-GGUF↗

  OsaurusAI

  **Diplodocus preview 4**, the Osaurus assistant for finance, documents, spreadsheets and tool-assisted work, built on IFM/K2-Horizon-7B and trained on Osaurus' own workspace tools (workbooks and documents edited in place, the built-in `calculate` tool, reasoning on and off). This is the **GGUF build for llama.cpp** in three calibrated sizes: Using Osaurus on a Mac? Use the MLX bundle `OsaurusAI/Diplodocus-preview-JANG_6M` instead. Same weights, same template, same reasoning toggle. **Requires llama.cpp with K2 Horizon support** (ggml-org/llama.cpp#29535, merged 2026-10-06). Older builds, and apps that bundle an older llama.cpp, will not load the `k2-horizon` architecture.

• TwIL-LM3↗

  webAI-Official

  A 3B reasoning model for **formal logic** tasks, built from `HuggingFaceTB/SmolLM3-3B` through LoRA supervised fine-tuning, checkpoint fusion, WiSE-FT weight interpolation, and entropy-weighted GRPO reinforcement learning. It improves in-domain formal-logic performance by **+26% relative** over its base model (macro gate 0.336 → 0.422) **and improves held-out benchmark performance at the same time** (+0.022 core average). It is the only arm in this project that gains on both tracks, which is why it is the recommended release of the pair. * **Gains on both tracks at once** — the only arm in this project that does. In-domain macro gate 0.336 → 0.422, and the held-out 10-dataset macro 0.7193 → 0.7339 rather than the usual collapse that follows task-specific fine-tuning. * **Beats every arm up to and including LFM2.5-8B-A1B** — roughly three times its parameter count — on all six Track A objective lanes and all four summary rows, not on average alone. * **Competitive with 8B on strict scoring.** On strict-7, which gives no loose-match credit anywhere, it sits 0.012 behind Qwen3-8B (0.1971 against 0.2093) at 2.6x fewer parameters, and ahead of it on Lean formalisation (token-F1 0.5869 against 0.4022) and semantic parsing (0.4416 against 0.4257). * **Structured formal output.** Tuned for the objects rather than the prose: FOL translation, entailment labels, semantic parses, Lean statements and Lean proof critique. * **The most efficient arm measured, at any scale.** 482-token Track B generations and 32.9 completed answers per second — about eight times gpt-oss-120b's rate — because it answers short rather than because it decodes unusually fast. * **Lowest maths-corpus perplexity of any released arm in the table** (3.8229), including Qwen3-8B at 4.0083. * **Runs anywhere.** 3.08B parameters in bf16, with Q4\_K\_M GGUF at 1.78 GiB for CPU or 4 GB of VRAM.

• mask2former-swin-large-coco-instance↗

  facebook

  Mask2Former model trained on COCO instance segmentation (large-sized version, Swin backbone). It was introduced in the paper Masked-attention Mask Transformer for Universal Image Segmentation and first released in this repository. Disclaimer: The team releasing Mask2Former did not write a model card for this model so this model card has been written by the Hugging Face team. Mask2Former addresses instance, semantic and panoptic segmentation with the same paradigm: by predicting a set of masks and corresponding labels. Hence, all 3 tasks are treated as if they were instance segmentation. Mask2Former outperforms the previous SOTA, MaskFormer both in terms of performance an efficiency by (i) replacing the pixel decoder with a more advanced multi-scale deformable attention Transformer, (ii) adopting a Transformer decoder with masked attention to boost performance without without introducing additional computation and (iii) improving training efficiency by calculating the loss on subsampled points instead of whole masks.

• leosamsHelloworldXL_helloworldXL60↗

  misri