保存済みツールを推薦
FOXAI SEARCH

サイト内検索

AI モデル 2119

  • SenseVoiceSmall-gguf↗

    handy-computer

    GGUF conversions of FunAudioLLM/SenseVoiceSmall for use with transcribe.cpp. Ported from upstream commit 3eb3b4e, pinned 2026-05-06. Validated against the FunASR reference at transcribe.cpp commit f094d28 on 2026-05-06. Offline multilingual speech-to-text in Chinese, Cantonese, English, Japanese, and Korean. A SAN-M encoder with a single CTC head over a 25,055-token SentencePiece vocabulary. Takes a 16 kHz mono WAV (capped at 30 seconds per call, per upstream's direct-inference contract) and produces a transcript. Not a streaming model, no translation, no built-in long-form chunking. The same CTC head also emits language-ID, simple emotion labels, audio-event tags, and inverse-text-normalization control tags. These tags are hidden unless `--raw-tokens` is passed. ITN is on by default for readable casing, punctuation, and digits; pass `--no-itn` for upstream's spoken-form output.

  • OpenThai-SystemOne↗

    iapp

    **An open Thai + English "System One" decision model.** It does not generate text. Given a *state* (any text or JSON) and typed *questions*, it returns calibrated probabilities over the options in **one forward pass**: The request/response contract mirrors TypeSafe's `POST /v1/systemone` so code written for the TypeSafe SDK can be pointed at this model unchanged. Typical uses: ticket routing, moderation, intent detection, RAG relevance judging, LLM-output verification, and **computer-use / browser-agent action selection** (which element to click, which tool to call). Ollama 0.35 serves System One models locally at `POST /v1/systemone`, with the same Jev-style contract. OpenThai-SystemOne has an **Ollama build**: `iapp/openthai-systemone` on ollama.com (tags `0.8b` = `0.8b-q8_0`, `0.8b-q4_K_M`, `0.8b-bf16`), also as GGUF on Hugging Face: `iapp/OpenThai-SystemOne-Ollama`. No API key; nothing leaves your machine.

  • SGuard-ContentFilter-2B-v1↗

    SamsungSDS-Research

  • Llama-3.1-Swallow-8B-Instruct-v0.5↗

    tokyotech-llm

    Llama 3.1 Swallow is a series of large language models (8B, 70B) that were built by continual pre-training on the Meta Llama 3.1 models. Llama 3.1 Swallow enhanced the Japanese language capabilities of the original Llama 3.1 while retaining the English language capabilities. We use approximately 200 billion tokens that were sampled from a large Japanese web corpus (Swallow Corpus Version 2), Japanese and English Wikipedia articles, and mathematical and coding contents, etc (see the Training Datasets section of the base model) for continual pre-training. The instruction-tuned models (Instruct) were built by supervised fine-tuning (SFT) on the synthetic data specially built for Japanese. See the Swallow Model Index section to find other model variants. **Note**: Llama-3.1-Swallow-8B-Instruct-v0.5 model was continually pre-trained from the meta-llama/Llama-3.1-8B-Instruct and then instruction-tuned with our instruction datasets. This release enhances the conversation capability of Llama 3.1 Swallow. The model is trained to imitate the behavior of gemma-3-27b-it. Among all open-source LLMs with <= 8 billion parameters, Llama-3.1-Swallow-8B-Instruct-v0.5 exhibits **state-of-the-art performance on Japanese MT-Bench**, outperforming its predecessor, Llama-3.1-Swallow-8B-Instruct-v0.3, by 1.5 points.

  • GLM-5.3-Flash-GGUF↗

    vcruz305

    Community llama.cpp GGUFs of Z.ai's GLM-5.3-Flash (`Glm5NextForConditionalGeneration`, 320B-A18B, native multimodal). **This pack is the first working GGUF conversion of GLM-5.3-Flash.** The convert registers GGUF arch `glm5next` and maps the 5.3 hybrid end-to-end: **KDA** (gated delta-net), **NoPE sparse MLA** + lightning indexer + **IndexPool**, **mHC**, **288-expert MoE** + shared expert, **MTP / NextN**, and the **vision tower as mmproj**. It is a 5.3 convert, not a GLM-5.2 `GlmMoeDsa` class rename (that path only covers the 11 MLA blocks and drops the 34 KDA layers). `llama-quantize` on that same `glm5next` build wrote the K-quants below (Q2_K 2.91 BPW, Q4_K_M 4.81 BPW). **Run** these GGUFs on vcruz305/llama.cpp `main` / `glm5next-mtp` (`ea37b8b`): Unsloth glm5next graph plus **NextN `graph_mtp`**, with the trunk emitting `t_h_nextn` so `--spec-type draft-mtp` can accept. ggml-org drafts #27752, #27754, #27773 load the weights; MTP decode is on this fork.

  • index-echo-9b-GGUF↗

    cstr

    Native speech translation with bilingual, timestamped subtitles, primarily Chinese → English, Japanese or Spanish. This repository contains the validated F16 encoder/connector and decoder pair for CrispASR. Keep both GGUF files in the same directory. Use a CrispASR build containing the 9B integration; earlier 2B-only builds cannot load this pair. `--target-lang` selects `en`, `ja` or `es`. Complete-file processing also uses the standard `ggml-silero-v6.2.0.bin` VAD companion; `--auto-download` fetches it. The 2B model remains CrispASR's smaller default. This 9B repository supports F16 only.

  • Qwen3.8-27B-int4-AutoRound↗

    Frozenlock

    INT4 (W4A16) AutoRound quantization of Qwen/Qwen3.8-27B — **18 GB**, with the **MTP speculative-decoding head quantized and working**. The MTP head is quantized along with the model, so `--speculative-config` in vLLM works out of the box and roughly **doubles decode throughput** (measured +98 % on temp-0 prose vs the same GPU without MTP; draft acceptance 41–67 % depending on content). At the time of upload, the available 4-bit options either dropped the MTP tensors (no speculative decoding) or kept hundreds of tensors in bf16 (28 GB — no longer fits a 24–32 GB GPU next to a useful KV cache). This one follows the recipe that served Qwen3.6-27B in production for months:

  • MiMo-V2.6-Flash-RL-cmf↗

    infosave

    > **Model release candidate.** All three weight files are uploaded and SHA-256 verified. > Text and greedy MTP checks pass; strict multimodal acceptance is still incomplete. > **Engine: Cortiq 0.7.7** · crates.io. Older published versions do not include this MiMo integration. **These are two complete file sets, not two copies of the 164-GB backbone.** Full includes the same text weights plus a **1.241-GB multimodal companion**. To upgrade from Text-only, download only the `.mm.cmf` file into the same folder. Text-only does not download or load the media towers. MiMo is a mixture-of-experts model with approximately **310 billion total parameters**, not a 15-billion-parameter checkpoint. Only a subset (roughly 15 billion parameters) is active per token, but **all experts must be stored**. Even a hypothetical uniform four-bit representation of 310 billion parameters would occupy about **155 GB**, before scales, metadata and higher-precision tensors. This conversion intentionally retains q8 and float tensors where needed.

  • MiniMax-H3-Qwen3-VL-NVFP4↗

    6block

    An **NVFP4 (4-bit) Qwen3-VL-32B text encoder** for MiniMax-H3, made with `convert-to-quant`. It is a drop-in replacement for the H3 conditioning text encoder (ComfyUI `CLIPLoader`, type `minimax`). The H3 video weights are unchanged — only the text encoder is quantized to 4-bit, to bring the full H3 stack onto consumer GPUs. This encoder deliberately keeps more of the network in bf16 than a minimum-size 4-bit encoder does. The `--qwen_vlm` filter protects — at bf16 — the token embeddings, the **first and last language layers**, and the **entire vision tower**, quantizing only the middle language-layer linears to NVFP4. The result is 16.64 GiB, about 2 GiB above the most aggressively quantized community NVFP4 encoders (≈14.6 GiB, e.g. the AWQ-quantized encoder in `Comfy-Org/MiniMax-H3`), in exchange for leaving the precision-sensitive layers (embeddings, boundary layers, vision) untouched. It is a fidelity-first tradeoff: spend ~2 GiB more VRAM to keep the layers that most affect conditioning at full precision. Everything still fits a 24 GB card with offload. Everything is held constant except encoder precision — the bf16 encoder vs this NVFP4 encoder, same prompts, same seeds, same DiT — over a 16-clip paired set.

  • qwen3.8-27b-w4a16-autoround-rtx3090↗

    grevinden

    Готовая квантизация **Qwen/Qwen3.8-27B** под одну карту **NVIDIA RTX 3090 24 ГБ** (Ampere, sm86). W4A16 int4, group_size=128, симметричный, формат compressed-tensors / **Marlin int4** — работает на инференс-стеке vLLM-форка `syv-ai/qwen38-27b-rtx3090` (и других vLLM с поддержкой compressed-tensors). Можно просто **стянуть и использовать** — без повторной квантизации. Это "серверный" вариант: логика `lm_head`/`embed_tokens` переквантована в int8, MTP-модуль квантизован, приложен draft-vocab для многотокенного предсказания. 1. **Калибровка** (стратифицированная, n=520, по приоритетам качества): | Источник | n | Цель | |---|---|---| | magicoder | 63 | Python-код, инструменты | | agentinstruct | 48 | агентный режим | | glaive_fc | 48 | function-calling / инструменты | | nemotron_terminal | 48 | работа с терминалом (файлы, отладка, данные) | | oasst2_ru | 48 | живой RU-чат, неформальное общение | | fineweb2-ru | 48 | русский текст | | pile-10k | 48 | общий EN | | dolly | 48 | резюме-анализ / качественные тексты | | gsm8k | 40 | рассуждения/математика | | 1c_code | 40 | 1С:Предприятие (BSL) | | ru_turbo_saiga | 26 | RU-инструкции | | humaneval | 15 | Python-код | Eval: n=84 из тех же источников (hold-out). 2. **AutoRound**: `--algs awq,signround --iters 800 --batch 2 --seqlen 2048 --nsamples 200 --seed 42` (рецепт против OOM Linux с гессианом — см. репозиторий-руководство). 3. **Подготовка под форк**: `lm_head`/`embed_tokens` → int8, MTP-модуль квантизован, сборка `mtp_draft_vocab_ids.pt` (40 960 токенов), 68 shards. 4. **Верификация** в контейнере инференса: `verify: OK (0 failures)`.