サイト内検索

AI モデル 2119

• DeepSeek-V4.1-Flash-1017-FP8-GGUF↗

  AMAImedia

  ⚡ Each donation funds the next large quant. I host free GGUF or MoE quants as independent research. Local hardware: Mechrevo Kuangshi GM7AG0M — RTX 3060 Laptop 6GB GDDR6, 64GB DDR5, i7-12700H (14C/20T, 4.7GHz), Windows 11, Samsung 990 Pro. Good for imatrix and 0.6–35B-class work in RAM. 9B+ and searches need rented H200/Blackwell , typically $100 per quant. Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators) **DeepSeek-V4.1-Flash-1017** is a minor update to **DeepSeek-V4.1-Flash**, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks. Alongside these gains, the model reduces task execution time and output length, enabling users to complete tasks faster and at lower cost and advancing the performance–cost Pareto frontier across multiple agent harnesses.

• Tristan_-_dclm-perplexity-correlations-spearmanr-no-samp-160m-gguf↗

  RichardErkhov

• Gemma-4-12B-a4B-MoE↗

  stamsam

• RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored-IQ3_S-Strata↗

  jimmy31chen

  A **Strata-ready** repackaging of `0bserverx/RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF` (IQ3_S quant family): 1. **`RVN-Qwen3.8-Flash-Next-IQ3_S-merged.gguf`** — the original 8-shard *split* GGUF merged into a **single file** (83 GiB), ready to feed directly to a Strata native pack. 2. **`pack/`** — a **pre-built Strata pack** (from the same version of `iq_pack.py --compat-bf16` shipped in Strata 0.1.30). Point `--pack` at it and Strata reads the experts/PLE table straight from the merged GGUF — no merge step, no pack step, no calibration needed for a first run. The upstream RVN release ships as 8 shards, and some transformer layers' expert tensors are **split across shard boundaries**. Strata's `tools/iq_pack.py` requires the gate/up/down expert tensors of each layer to live in the *same* file, so the upstream files fail to pack (`layer 1: its gate/up/down tensors are in different shards`). This repo provides the bit-identical merge so Strata users can use RVN out of the box. Experts layout: gate/up = `IQ3_S`, down = `IQ4_NL` (both natively supported by Strata's MMQ kernels). PLE n-gram table = `IQ4_NL`, read directly from the merged GGUF.

• MiniMax-M3.1-Flash-Maya-GGUF↗

  peasantsmith

  GGUF quantizations of MiniMaxAI/MiniMax-M3.1-Flash (428B parameters, 23B active per token: 128 routed experts per layer, 4 active, plus a shared expert; sparse attention in every layer; up to a 1M-token context), made for **Project Maya**. Project Maya keeps the most-used experts on the GPU and computes the rest from RAM, so the model runs well below its size. Both quants are made from MiniMax's own release (FP4 experts, FP8 dense weights), not from a re-quantized file, and keep every expert. Measured on held-out conversations never used for calibration, on the model's own answers, against the original model's predictions (MiniMax's released weights, decoded exactly), with the files as published running in Project Maya's engine - the same results on NVIDIA and AMD. 1. **Calibration text** in MiniMax's own chat format, thinking included: tool calls and agent traces, code (with three.js / canvas / WebGL and other front-end work), reasoning, long answers, chat, multilingual chat - 384 sequences of 2,048 tokens. A separate held-out set is kept for evaluation. 2. **Statistics from the model itself**, layer by layer: an importance matrix for every expert separately. 3. **Where the bits go:** every layer's gate, up and down matrices were measured in each candidate type, and the extra bits went to the layers where the model is most sensitive. 4. **Error-feedback rounding** of every expert, at the quantizer's 256-weight blocks, each expert against its own input statistics - and each layer calibrated on what the already-quantized layers before it produce. 5. **llama.cpp's own quantizers (ggml)**, so the files are ordinary GGUFs in llama.cpp's `minimax-m3` layout.

• parrot_paraphraser_on_T5↗

  prithivida

  Parrot is a paraphrase based utterance augmentation framework purpose built to accelerate training NLU models. A paraphrase framework is more than just a paraphrasing model. For more details on the library and usage please refer to the github page **Huggingface** lists 12 paraphrase models, **RapidAPI** lists 7 fremium and commercial paraphrasers like QuillBot, Rasa has discussed an experimental paraphraser for augmenting text data here, Sentence-transfomers offers a paraphrase mining utility and NLPAug offers word level augmentation with a PPDB (a multi-million paraphrase database). While these attempts at paraphrasing are great, there are still some gaps and paraphrasing is NOT yet a mainstream option for text augmentation in building NLU models....Parrot is a humble attempt to fill some of these gaps. **What is a good paraphrase?** Almost all conditioned text generation models are validated on 2 factors, (1) if the generated text conveys the same meaning as the original context (Adequacy) (2) if the text is fluent / grammatically correct english (Fluency). For instance Neural Machine Translation outputs are tested for Adequacy and Fluency. But a good paraphrase should be adequate and fluent while being as different as possible on the surface lexical form. With respect to this definition, the **3 key metrics** that measures the quality of paraphrases are: - **Adequacy** (Is the meaning preserved adequately?) - **Fluency** (Is the paraphrase fluent English?) - **Diversity (Lexical / Phrasal / Syntactical)** (How much has the paraphrase changed the original sentence?)

• qwen3-4b-mamba3-hybrid-stage2b-kd-bias0↗

  arianraje

  **Stage-2b endpoint, bias-0 arm** -- long-context forward-KL logit distillation against Qwen3-4B at sequence length 32,768, 294M tokens (141 steps of 2,097,152; micro-batch 1 x 4 ranks x 16 accumulation), from the stage-2a endpoint of this arm. LR halved from stage 2a: mixers 1e-4, inherited weights 1e-5, cosine to 0.1x with 3% warmup; all weights train. 4x B200. Final val KL at 32k **0.2021**, top1-agreement **0.8056**, wikitext-2 PPL (4096 context) **11.93**, from a starting point of 1.1727 / 0.7167 / 12.49. At this same stage the Qwen GDN arm reaches KL 0.2177 / top1 0.7985 and the Qwen Mamba2 arm 0.2336 / 0.7875; see the table below for the bias-1 arm. This is the input to stage-3 on-policy distillation, which has not been run. No downstream evaluation yet. This is a **custom container** (`model_type: qwen3_mamba3`), not a stock architecture, so `trust_remote_code=True` is required. The modeling code is bundled in this repo (`modeling_qwen3_mamba3.py`) and needs no other checkout:

• qwen3-forced-aligner-0.6b↗

  OpenASR

  **Word-level forced alignment for OpenASR transcripts -- a non-autoregressive Qwen3 audio+text model that refines per-word timestamps** A capability-pack support model for the **OpenASR** runtime — pure-Rust inference, **no Python at inference time**. Not a standalone transcription model: it augments another OpenASR ASR model's own decode path. Qwen3-ForcedAligner-0.6B is a **word-level forced-alignment** model from **Qwen**, sharing its audio-encoder + LM `thinker` tensor layout byte-for-byte with **Qwen3-ASR** (same `Qwen3ASRForConditionalGeneration` architecture). The only structural difference is the final head: instead of a tied vocabulary `lm_head`, it uses an independent `Linear(hidden_size, 5000)` classification head over 80ms-wide timestamp bins. Given a transcript's text and its source audio, it runs a single non-autoregressive forward pass and reads off word-boundary timestamps at argmax ` ` positions -- refining a model family's own (typically decode-time-approximate) per-word timestamps. This OpenASR repo repackages the weights as `.oasr` packs that run natively in the OpenASR runtime -- no Python at inference, all decoding local. OpenASR recommends the validated **q4_k** tier and also ships **q8_0** and an **fp16** full-precision reference tier. The q4_k label is the catalog's unified product name, not a claim that every matrix uses Q4_K: the audio encoder, token embedding, and timestamp head remain Q8_0, and pack verification replays the exact per-tensor policy. Q3 and legacy all-Q4 packs are rejected because small logit perturbations can move a word boundary across multiple 80ms bins.

• Kimi-VL-A3B-Thinking↗

  月之暗面 · Kimi

  > [!Warning] > This model has a new version: Kimi-VL-A3B-Thinking-2506. Please consider using the new 2506 version for better abilties on general visual understanding, reasoning, video and agent scenarios. *Please set a higher temperature for thinking model, especially when your problem requires relatively long thinking process. We have updated the following sample script for HF inference.* We present **Kimi-VL**, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers **advanced multimodal reasoning, long-context understanding, and strong agent capabilities**—all while activating only **2.8B** parameters in its language decoder (Kimi-VL-A3B).

• LFM2.5-1.2B-JP-202606-ONNX↗

  LiquidAI

  ONNX export of **LFM2.5-1.2B-JP-202606** for cross-platform deployment via ONNX Runtime, Transformers.js, and the WebGPU stack. Same weights, same chat template — just compiled into ONNX graphs at multiple precisions, including a WebGPU-friendly INT4 + FP16 mix. LiquidAI/LFM2.5-1.2B-JP-202606 は当社の汎用日本語チャットモデルです。本リポジトリはその ONNX エクスポートで、ONNX Runtime / Transformers.js / WebGPU での実行に対応しています。重みおよびチャットテンプレートは同一です。 > **`model_q4f16.onnx` is the recommended variant for WebGPU**: INT4 weights with FP16 scales, FP16 KV cache and conv state I/O, FP32 logits via an inserted Cast — the format Transformers.js targets for browser inference.