뉴스 및 전체 검색
뉴스, 도구, 모델, 가이드, 주제, 워크플로, 프롬프트를 함께 검색합니다.
AI 모델 34
“The Full Stack”
MOSS-Audio-Tokenizer-Nano
This repository contains the Hugging Face remote-code implementation and weights for **MOSS-Audio-Tokenizer-Nano**, the lightweight audio tokenizer used by **MOSS-TTS-Nano**. MOSS-Audio-Tokenizer-Nano is a compact discrete audio tokenizer based on the **Cat** (**C**ausal **A**udio **T**okenizer with **T**ransformer) architecture from MOSS-Audio-Tokenizer: Scaling Audio Tokenizers for Future Audio Foundation Models. The checkpoint in this repository has **21,969,664 parameters** (approximately **22M**), making it much smaller than the full-size MOSS-Audio-Tokenizer while preserving the 48 kHz stereo tokenizer interface used by the MOSS-TTS family. - **Small model size**: approximately **22M parameters**, including about 10.45M encoder parameters, 10.45M decoder parameters, and 1.07M quantizer parameters. - **Native high-resolution audio**: supports **48 kHz** input and output with **2-channel stereo** audio, helping reduce compression loss and improve listening quality. - **Low-frame-rate discrete codes**: compresses 48 kHz stereo audio into a **12.5 Hz** token stream with a downsample rate of 7,680 samples. - **Variable bitrate reconstruction**: uses a residual quantizer stack with **16 codebooks** and 1,024 entries per codebook. Each codebook contributes about **0.125 kbps**, for an inference range from **0.125 kbps to 2 kbps**. - **Transformer-based tokenizer**: uses causal Transformer blocks and supports low-latency streaming encode/decode. - **MOSS-TTS family interface**: designed as the audio tokenizer backbone for MOSS-TTS-Nano and compatible MOSS-TTS-family workflows.
tomoro-colqwen3-embed-4b
**TomoroAI/tomoro-colqwen3-embed-4b** is a state-of-the-art ColPali-style multimodal embedding model. It maps text queries, visual documents (images, PDFs) or short videos into aligned multi-vector embeddings. Built by merging **Qwen/Qwen3-VL-4B-Instruct** with **Qwen/Qwen3-Embedding-4B**, this model inherits robust text retrieval capabilities while preserving a full vision stack. It has been fine-tuned on a curated mixture of VDR, ViDoRe-ColPali-Training, VisRAG-Ret-Train-Synthetic-data, and VisRAG-Ret-Train-In-domain-data. It achieves SOTA or competitive performance across **ViDoRe V1-V3** (English and Multilingual) while offering a significantly reduced embedding footprint compared to other full-dim Colpali model alternatives. ColQwen3 generalizes to short videos while learning from image-text retrieval task. This minimal example samples a clip with `torchvision`, encodes queries and frames, then pools frame embeddings with a per-dimension max before MaxSim scoring.
NVIDIA-NemotronLabs-VoiceChat-11B
NVIDIA NemotronLabs VoiceChat is a **11B end-to-end, real-time speech full duplex (FD)** model for conversational AI that jointly performs streaming speech understanding and speech generation [1, 2, 3]. Unlike traditional cascaded stacks (ASR → LLM → TTS), this model achieves full duplex, real-time, seamless voice interaction in one unified architecture, eliminating the need for multiple models or API handoffs, thus reducing end-to-end latency. It sets new benchmarks by bringing open, robust, and highly natural conversation capabilities. Moreover, NVIDIA NemotronLabs VoiceChat is **the first open full-duplex model to support tool calling** while maintaining a natural conversation flow during tool execution. For each tool, a specific “on-hold” message can be defined that will be spoken by the agent as soon as the LLM generates the text that will trigger the tool call and response. The model operates on audio signals, which are encoded using a fast conformer module. The resulting audio tokens are inputted into a Nemotron Nano V2 9B LLM backbone to predict text tokens, which are fed to a TTS decoder to predict audio codes for generating the agent's speech. A separate output channel is used to predict tool calling scripts. NemotronLabs VoiceChat offers an unprecedented trade-off between intelligence and latency in the space of open-source voice agents, as highlighted by our benchmarking results below. NemotronLabs VoiceChat is targeted for researchers, developers, and professionals in the field of natural language processing (NLP) and speech technology for purposes such as automatic speech recognition (ASR), text-to-speech synthesis (TTS), and voice assistant development.
Qwen3.8-27B-nvfp4qat-NInfer
The repository contains a QAT-sourced NVFP4 weight profile of Qwen/Qwen3.8-27B in the native NInfer `.ninfer` artifact format, with the z-lab/Qwen3.8-27B-DFlash2 block-diffusion speculative drafter embedded in the same image. It runs on NInfer **v3** engines — Text and MTP components through format/shape routes the upstream runtime already implements, the optional NVFP4-encoded DFlash2 companion through the fork's runtime support. It is not a Transformers checkpoint, Safetensors distribution, or GGUF file. This is a fourth weight profile for the existing `qwen3_8_27b` target — a peer of the official `groupwise-int` and `nvfp4` profiles and of the fork's fuller-requant `nvfp4full` profile. The Text weight stack is copied **word-for-word** from QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 — a quantization-aware-trained checkpoint (QUASAR loss-aware NVFP4 distillation against the frozen BF16 teacher, arXiv 2608.13966): - every one of the 496 text linear layers is NVFP4 (W4A4) — attention, gated-delta-net, and MLP alike, with no high-precision exceptions; - the QAT factory quantizes per site, so every constituent tensor of a fused parent shares one weight and input global scale; the converter enforces that sharing before copying a word, and the site input divisors ride the artifact directly; - the GDN control projections are decoded from their QAT NVFP4 words to the BF16 control parent the engine consumes.
Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16
> [!IMPORTANT] > **Superseded by Qwen3.8 — use this instead:** `AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4-MIXED` > > **Operator recipes (Spark Dynamic DFlash lattice, TP=2, RTX):** `AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED` > > This Qwen3.6-27B card stays online for history and existing downloads. For new work, move to the Qwen3.8 Ultimate Uncensored NVFP4-MIXED release. That line is far superior in capability and reaches substantially more of what the architecture can do, with a more refined mixed quantization methodology (NVFP4 + FP8 lattice) than these older Qwen3.6 exports. > > Serve with the unified container: `ghcr.io/aeon-7/aeon-vllm-ultimate:latest` (Spark / GB10) or `ghcr.io/aeon-7/aeon-vllm-ultimate-rtx:latest` (RTX discrete). > **Deployment, operations & benchmarks → github.com/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-DFlash** > > The GitHub repo is the source of truth for the production deployment guide, hardware-tuned docker-compose configs (DGX Spark NVFP4, A100/H100 BF16), full configuration reference, measured throughput benchmarks, and `AGENTS.md` — an operator's manual that pre-empts common stale-documentation traps for AI coding agents working on this stack. > **🆕 2026-05-01 — MTP head grafted in.** This repo now ships with the original `mtp.*` head (15 tensors, ~0.85 GB) restored from the `Qwen/Qwen3.6-27B` base. vLLM's `--speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":3}'` works directly on the BF16 checkpoint with no extra steps. Measured on DGX Spark: **mean accepted length 3.3/3, P0 ≈ 90% acceptance, avg draft acceptance 78%** — comparable to the base model, confirming abliteration doesn't damage the model's top-K distribution that MTP relies on. Credit to `@tcclaviger` for the empirical finding (discussion #6) that MTP doesn't need retraining post-abliteration. **No retraining was performed.** The MTP head is the unmodified base — only the residual-stream-writing weights of the LM were orthogonalized in the original v8 abliteration pass; the MTP head is independent of that pathway.
Qwen3.6-27B-Abliterated-Heretic-Uncensored-GGUF
This is a GGUF release of an abliterated, uncensored version of Qwen's Qwen3.6-27B with vision, made with Heretic. By applying a Heretic-style two-stage MPOA pipeline with magnitude preservation on the Qwen3.6-27B dense text stack — slot-grouped output-side ablation followed by jailbreak-conditioned input-side ablation, mirroring the methodology used for Qwen3.6-35B-A3B — the base refusal behavior was attenuated at the weight level with low distributional divergence (KL 0.0282 vs base on harmless prompts). The result keeps Qwen3.6-27B's full vision and video multimodal architecture and general capability profile intact, while no longer defaulting to the original refusal pattern. Numbers measured by reading every response, not by regex / refusal-marker scoring. Greedy decoding (`do_sample=False`), `enable_thinking=False`, on `mlabonne/harmful_behaviors test[:25]`. KL on `mlabonne/harmless_alpaca test[:25]`. The 3 remaining deflections under the jailbreak system prompt are crisis-substitution patterns (violent crime → conflict resolution, suicide → crisis hotline, car theft → legal vehicle acquisition).
Ling-3.0-flash-abliterated-APEX-GGUF
Ling-3.0-flash-abliterated-APEX-GGUF English | 📖 中文文档 inclusionAI Ling-3.0-flash · 124B-A5.1B hybrid MoE · abliterix v6 trial 21 · APEX GGUF Ling-3.0-flash is inclusionAI's next-generation native hybrid reasoning model: 124B total / 5.1B active (~12.4% of their previous 1T-class flagship). It uses a native hybrid-linear stack from pretraining — 5:1 Kimi Delta Attention (KDA) + gated MLA , 1/64 sparse MoE, 512 routed experts (top-8) + 1 shared expert, 2 dense layers, and a trained MTP head. Official context schedule is 8K → 32K → 256K. Thinking is on by default in the official card. This community package is not an official inclusionAI release . It applies two extra steps on top of the official weights: Refusal edit (mergeable recipe) via abliterix v6 trial 21 LoRA, stream-merged back to BF16. Write path: MPOA weight_normalization=full , LoRA rank 3, o_proj + down_proj only. This is the static merge, not the later runtime-only gate (v36). GGUF : BF16 convert + APEX mixed-precision tensor-type files (no imatrix on these tiers). Architecture in the GGUF is bailingmoe3 .
JiRackUltra_32b
A fast and efficient 32B model optimized for CPU inference. The model was refactored with BitNet features and an updated tokenizer that includes new **Routing**, **Media**, **Vision**, **Sound**, **Tool call**, and **Robotics** tags. Built on a DeepSeek R1-32B architecture with native ternary (BitNet-style) support and ready-to-run GGUF quantizations. - JiRack is a cloud-ready model that helps save money on cloud infrastructure. It can be used as an expert model in RAG deployments, with the ONNX JiRack Java server as an alternative. Once the container is running, open your browser and navigate to: `http://localhost:7869` This opens the JiRack UI — a clean web interface. The listening port can be easily modified directly from the **Settings** panel within the JiRack UI. - Model weights are released under the MIT License — free to use, modify, and distribute for any purpose, including commercial. No royalties, no per-user fees, no subscription. - The Docker image with UI and the pre-built Ollama quantizations are separate paid products. If you prefer to build your own secure deployment — take the weights, assemble your own stack, and you're done. - The JiRack Ultra 32B model for Docker and Ollama is provided under a commercial license ($12 per user per year). - All JiRack UI clients are provided under a commercial license. - However, the UI clients can be used for free when running together with the official JiRack Docker containers, as long as they are not redistributed separately. For commercial licensing, cluster deployment, or enterprise use of JiRack models, please contact us. - **JiRack MS Windows 11 Desktop Client (with Ollama API):** https://huggingface.co/kgrabko/JiRackTernary_1b/resolve/main/jirack-chat.zip - **Live email chat with the model:** support@cmsmanhattan.com | Use Case | CPU | RAM | Recommended Quant | Expected Speed | Recommendation | |-------------------|----------------------------|-----------|-------------------|---------------------|----------------| | Recommended | Ryzen 9 / Intel i9 / Xeon | 32–48 GB | Q4_K_M | Good interactive | Best choice | | High Performance | High-core server CPU | 64 GB+ | Full / Q4 | Excellent | Excellent | | Low
Qwen3.8-Flash-Next-MTP-GGUF
> **20 Sep 2026 — new default is Q5_K, not Q8_0.** A lighter draft head turned out to > be faster *and* to get accepted more often: +2.8 % mean across six workloads, +8.7 % > on prose. Grab `mtp-Qwen3.8-Flash-Next-Q5_K-frspec-65k.gguf` under Files. Q8_0 stays > up and still works. Numbers under What it buys, and the reason the > file is requantised rather than built clean is under > Provenance — it measured better that way, which surprised me too. > > **Still Vulkan, not ROCm** (17 Sep): ROCm returns wrong answers on this chip without > telling you. The old ROCm table is at the bottom. A small head that guesses the next few tokens so the big model can verify several at once. Useless on its own — it only runs alongside the full model. Part of an ongoing project to optimise Qwen3.8-Flash on Strix Halo. Some of the dev work is my own, some builds on community releases. When I have time I try to merge upstream. Full stack and numbers.
MOSS-Music-8B-Instruct-GGUF
GGUF conversion of OpenMOSS-Team/MOSS-Music-8B-Instruct for **HOT-Step CPP**, a fully local desktop app for AI music generation with a native C++/GGML engine. As far as we know this is the first GGUF conversion of this model. Unlike everything else HOT-Step ships, this model **analyses** audio rather than generating it. It listens to a track and writes a caption describing what it hears — used in HOT-Step's Training Studio to caption your own dataset locally, with no API key and nothing leaving the machine. These files are downloaded automatically by HOT-Step's Model Manager. They are **not** usable with llama.cpp alone: `moss-lm-*.gguf` is structurally a Qwen3 GGUF, but captioning requires the audio tower in `moss-aud-f16.gguf` and the deepstack injection wiring implemented in HOT-Step's engine.