저장한 도구도구 추천
FOXAI SEARCH

전체 검색

AI 모델 2119

  • Sage-A1.5-Sangreal-ICE↗

    gbuzhf

    GGUF quantizations of InternScience/Sage-A1.5, made with a Sangreal 10k importance matrix, ICE 1.6 per-tensor recipes and ICE 2.0 post-quantization expert repair. Every file keeps the MTP draft layer; the vision projector ships separately (InternScience's F16, plus a Q8_0). The ICE files carry the Arrow 4.0-sage chat template. Score: F16 = 100; 70 % full-distribution fidelity `1/(1+KLD)` + 30 % same-top-1, averaged over the three sets. The MTP head is in every file; on a llama.cpp build with MTP support, enable it for speculative decoding (draft depth 1). A Q8_0 of the projector (`mmproj-Sage-A1.5-Q8_0.gguf`) can be used in place of the F16 one.

  • gemma-3-27b-it-abliterated-GGUF↗

    mlabonne

    This is an uncensored version of google/gemma-3-27b-it created with a new abliteration technique. See this article to know more about abliteration. I was playing with model weights and noticed that Gemma 3 was much more resilient to abliteration than other models like Qwen 2.5. I experimented with a few recipes to remove refusals while preserving most of the model capabilities. In the original technique, a refusal direction is computed by comparing the residual streams between target (harmful) and baseline (harmless) samples.

  • vulnerability-severity-classification-roberta-base↗

    CIRCL

    The model was presented in the paper VLAI: A RoBERTa-Based Model for Automated Vulnerability Severity Classification [arXiv]. **Abstract:** VLAI is a transformer-based model that predicts software vulnerability severity levels directly from text descriptions. Built on RoBERTa, VLAI is fine-tuned on over 600,000 real-world vulnerabilities and achieves over 82% accuracy in predicting severity categories, enabling faster and more consistent triage ahead of manual CVSS scoring. The model and dataset are open-source and integrated into the Vulnerability-Lookup service. > **Note (October 2026):** since VulnTrain 3.4.0, > the dataset is split so that records of the same vulnerability (for example a CVE and its > GitHub advisory) never appear in both train and test. Accuracy measured this way (**74%**) > reflects performance on vulnerabilities the model has never seen. Earlier figures, including > the one in the abstract above, were measured on a split where many test descriptions also > appeared in training. Details in the technical report > Improving vulnerability-scores: Leakage-Free Splits and Consistent CVSS 4.0 Scoring.

  • Qwen3.8-Flash-Next-NVFP4-GPTQ-Strata↗

    Maximilian228

    Qwen's original Qwen3.8-Flash-Next with **every routed expert in NVFP4, quantized from the BF16 checkpoint by GPTQ**, packed for Strata NVFP4: the 125B hybrid MoE on one RTX 20-50 card (12 GB of VRAM or more) and 64 GB of RAM or more. - **Accuracy:** how often the model's most likely next token is the Q8_0 model's. Measured on 5,153 tokens of code, docs, English and Ukrainian chats and agent transcripts, every model fed the same text. Every model ran with its experts on the CPU and the route tail skip off, so only the quantization differs. The Q8_0 model also has Q8_0 attention, Qwen's BF16 router and Qwen's own FP8 PLE, so it is the original's precision everywhere but the 8-bit weights. - **Speed:** RTX 5090 (32 GB), Ryzen 9 9950X3D, 128 GB of DDR5-5600, Windows 11, Strata NVFP4 0.1.42-nvfp4.1's engine with the arguments 0.1.42-nvfp4.2's setup writes for each model. Median of 3 runs: 4 greedy chats (up to 1,500 tokens each) and one 32,007-token prompt. - **RAM:** "RAM in use" is the engine and the server at the speed shown; "PC for this speed" is the installed RAM that holds it with Windows beside it. Smaller PCs still run the models, slower: the NVFP4 GPTQ and NVIDIA's on a 64 GB PC in the low-RAM mode decode ~115 tokens/s and read a 32K prompt in ~9 s (docs/NVFP4.md, "64 GB of RAM"); Unsloth's files go into a RAM-budget mode there and read experts from the SSD (UD-Q4_K_XL: 7-8.5 tokens/s at 64 GB, upstream's measurement); setup's minimum is 48 GB for them and the ISTA 2-bit files, 60-64 GB for the ISTA 3-bit ones. - **Unsloth's files keep every expert in RAM where they fit with 16 GiB more** (setup from 0.1.42-nvfp4.2): from ~88 GB of RAM for UD-Q4_K_XL and ~68 GB for UD-Q3_K_XL. Below that, setup's RAM-budget mode reads the rest from the GGUF on the SSD. Setup before 0.1.42-nvfp4.2 used the budget mode on every PC; on 128 GB UD-Q4_K_XL then decoded 118 tokens/s and read the 32K prompt in 30 s (run setup again to switch an existing install). - UD-Q3_K_XL's experts are i-quants despite the name: IQ3_XXS gate/up with IQ4_NL or Q8_0 down. - The GGUF models run upstream's route tail skip by default: ~10% faster decode, and answers a little further from Q8_0 than the table shows. - The two 128 GB models need a 128 GB PC (setup asks below that). Q8_0 runs there in the low-RAM mode by itself (100 GiB of experts locked in RAM); the mix keeps every expert in RAM.

  • Carnice-Qwen3.6-MoE-35B-A3B-APEX-GGUF↗

    mudler

    ⚡ Each donation = another big MoE quantized I host 25+ 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 samuelcardillo/Carnice-Qwen3.6-MoE-35B-A3B. APEX is a quantization strategy for Mixture-of-Experts (MoE) models. It classifies tensors by role (routed expert, shared expert, attention) and applies a layer-wise precision gradient — edge layers get higher precision, middle layers get more aggressive compression. I-variants use diverse imatrix calibration (chat, code, reasoning, tool-calling, agentic traces, Wikipedia).

  • VibeThinker-3B-gguf↗

    samuelchristlie

  • StartLux-Decision-v1-27B-Q4_K_M-GGUF↗

    startlux-models

    StartLux-Decision-v1-27B in Q4_K_M (4-bit) as a GGUF file for llama.cpp. The decision server in this repository turns it into the same typed decisions, with a probability for every option, as the original model. "Same decision" is the share of the 231 public JevBench items on which the file picks the same answer as the original weights run through the `startlux_decision` package, with the same prompts, readout and temperatures. Q8_0 and Q4_K_M are llama.cpp's standard quantizations of BF16, without an importance matrix. `-ngl 99` puts every layer on the GPU (CUDA or Metal); leave it out on a CPU-only machine. llama.cpp has to be recent enough for this model (build b10454 or newer).

  • Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-DFlash2-NInfer-v3↗

    WaveCut

    A single-file NInfer **v3** artifact carrying the community uncensored tune HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF as the target model and the official z-lab/Qwen3.8-27B-DFlash2 draft model as the DFlash2 speculative companion (recipe `qwen3_8_27b-v2`, the first NInfer recipe with the DFlash2 payload). The native NextN/MTP head of the tune is preserved as well, so both `--spec mtp` and `--spec dflash2` work. It runs on **NInfer-all**, the `master` branch of iamwavecut/ninfer-all, which serves the RTX 3090, RTX 4090, RTX 5090 and RTX PRO 6000 Blackwell. Its first (v2) release ran on stock NInfer, which builds only for the RTX 5090; this v3 file was verified on NInfer-all. `--spec mtp --draft-tokens 3 --lm-head-draft` selects the tune's own NextN head instead. Vision (`--vision`) is the official tower: the release's BF16 projector reproduces it bit-for-bit through the FP16 export path, so the artifact carries the official BF16 vision words. **Hardware.** Upstream NInfer targets one RTX 5090 (`sm_120a`, 32 GB). NInfer-all serves the same `qwen3.8-27b / groupwise-int` identity with DFlash2 on the RTX 3090, 4090, 5090 and RTX PRO 6000: its Docker image needs an NVIDIA driver of the CUDA 13 branch (580 or newer) and the NVIDIA Container Toolkit, and a build (its README) sets `CMAKE_CUDA_ARCHITECTURES` to `86`, `89` or `120a`. The speed of the official artifact of this identity on each card is in its reference measurements, this tune itself was not measured there. On 24 GB cards the official artifact of the same size fits DFlash2 with five drafts beside 258,048 tokens of `rk4v4` KV on the 3090 and 245,760 on the 4090 (`--kv-dtype rk4v4 --gdn-state-fp16 --spec dflash2 --draft-tokens 5`).

  • LongLive-2.0-5B-Diffusers↗

    Rabinovich

    A diffusers-directory-layout repackaging of Efficient-Large-Model/LongLive-2.0-5B so it loads directly in **SGLang Diffusion** (`sglang.multimodal_gen`) without any runtime overlay/materialization. - **transformer/** — the `generator` weights extracted from the original `model_bf16.pt`, kept in their original (`model.*`) naming; SGLang's `LongLive2Transformer3DModel.param_names_mapping` maps them to the diffusers module names at load (same convention as LingBot-World). - **scheduler / text_encoder / tokenizer / vae** — taken from `Wan-AI/Wan2.2-TI2V-5B-Diffusers`. - `model_index.json` `_class_name = LongLive2Pipeline`. Original model & method: NVlabs/LongLive. License: NVIDIA Open Model License

  • piper-plus-tsukuyomi-chan↗

    ayousanz

    The zs-v1 release is the existing 500 epoch Tsukuyomi-chan fine-tune from the same v7 epoch32 generation as the public base. Use its ONNX and config together. This fixed voice requires no reference WAV or speaker encoder. The release includes a fine-tuning checkpoint, exact hashes, synthetic samples, credits and observed training provenance. It is an experimental baseline. Listening review status: **recorded**. Fresh three-text CAM++ mean SECS was 0.6617 (FT), 0.6538 (legacy) and 0.6956 (public-base Zero-Shot); these results do not establish improved voice quality. See the release card and verification record for the review scope and limits. The original root model files, configuration, samples and description below remain available for existing consumers and package aliases. Downloading through an older package alias still selects those original files.