Open-Weight Models

Files that the Buddy macOS app (an open-source screen buddy) bundles inside the app at build time. The build downloads them from a fixed revision of this repo and checks each file's SHA-256, so the app never downloads anything at run time. Each file keeps the license of the work it comes from. They are listed per folder below. What changed from the original model: the weights are the same model, stored in a different format. The Hugging Face safetensors weights were converted to ggml with whisper.cpp's conversion script and quantized to 8-bit (q80), which makes the file smaller (82 MB). Nothing was retrained or fine-tuned. How Buddy runs it: whisper-server -l hi -nt -bs 1 -bo 1 -t 4, on…

A small, fast network for Hivemind, a UCI engine for Bughouse chess — the four-player variant played on two boards, where captured pieces are passed to your partner. The network drives a Monte Carlo Graph Search (MCGS) that evaluates both boards jointly. It was distilled from the 13.6 M-parameter RISEv3.3 network and then improved by self-play reinforcement learning. At 0.97 M parameters it searches about 3x as many nodes per second, and was stronger at 100 ms per move in our matches. A small network that runs the same trunk on each board separately, with shared weights, and lets the boards exchange information explicitly. RISE-style networks stack both boards as channels of one 8x8 image…

Networks for Hivemind, a UCI engine for Bughouse chess, trained to play on real clocks. The TwinBoard-S networks learned bughouse with a single time-advantage bit: one team may sit (wait instead of moving) for the whole game. This lineage continues from the last of them and learns from self-play on a simulated clock instead, where a team may sit only while it actually leads, and every action it takes while on turn on both boards costs it its own move time. Same TwinBoardNet (twin-s, no attention) as the TwinBoard-S networks, reading 39 planes per board (78 in all): the 37 planes of input representation version 3, plus two clock planes that both boards carry with the same values, from the…

Model

bulmax-asm-baseline

Dakuwon Moody

A 2,067,130,880-parameter decoder-only transformer whose entire forward pass, backward pass and optimizer step run on hand-written x86-64 NASM kernels (AVX2+FMA), driven by a pure C runtime with no Python in the training loop. complex/quantum attention, no low-rank factorization. It exists as a clean CPU control alongside the main BulmaX multimodal run, on the same tokenizer (bulmasp.model) for data comparability. Raw binary BULMAXCP v1 -- not safetensors, not pickle. Fixed 64-byte header, 64-byte directory entries, 64-byte-aligned float32 blobs, so a C or assembly loader can mmap it and walk the directory with no JSON parser and no Python. Files hold weights plus both AdamW moments and the…

Model

bulmax-gen2-weights

Dakuwon Moody

BulmaX Gen2 is a custom generative multimodal research model with a language backbone, audio and vision pathways, a persistent learned coherence graph, and adaptive model-growth systems. This repository is a public weights and continuation-state mirror. It does not contain the private architecture or training source and is not a drop-in Transformers model. BulmaX's graph-context evaluation reached 5,242,880 tokens (the 5M stage), with all registered stages passing across three seeds. The design combines short-range neural attention with long-range coherence memory: the local attention window stays at 1,024 tokens, while context is streamed through 5,120 chunks with isolated per-document…

ggml conversions of BuzzASR/korean, a Korean fine-tune of openai/whisper-large-v3, for use with CrispASR and other whisper.cpp-compatible runtimes. All credit for the model goes to the BuzzASR authors — see the original model card for training details and benchmarks (FLEURS + Common Voice combined CER 4.61%, vs 5.72% for Whisper-large-v3 zero-shot). Always pass -l ko. The model is fine-tuned for Korean only. MIT, inherited from BuzzASR/korean. If you use this model, please cite the BuzzASR paper: arXiv:2609.09554 (Findings of EMNLP 2026).

Model

buzzasr-multilingual-demo

Christian Hoang

This repository contains the reviewed source for a Gradio demo covering 102 public BuzzASR automatic speech recognition models. It contains application code and sample audio; it is not a running Hugging Face Space and contains no model checkpoints. The app offers a language picker, loads the selected model on demand, preserves model-specific generation settings, handles up to five minutes of chunked audio, displays right-to-left transcripts where appropriate, and keeps at most two checkpoints in its dedicated local cache. Use Python 3.12 with an isolated environment: The included Persian examples use clips from the FLEURS fair test set under CC-BY. models.json lists the 102 public BuzzASR…

LoRA judge adapters for the stock JetLM/SDAR-30B-A3B-Chat-b32 @ c351bbc (block 32), trained with code/corpus1b/trainc1b.py from github.com/NoviceCoderInfinity/diffusion-moe-general-expert (branch phase1-datacorpus). The drafter they judge is the stock JetLM/SDAR-1.7B-Chat-b32 @ 2c83c05. Every folder / / is a complete checkpoint: adapter.safetensors + adapterconfig.json (LoRA r32 / alpha 64 on q, k, v, o in all 48 layers + the draft flag) and trainerstate.pt (AdamW state, token count) + trainerstate.json, so it can be both used and resumed (--resume / --continue-from). /config.json is the run's full configuration. The adapter is flag-gated: load it with draftalign.adapter.loadadapter and…

La primera instalación la hace un técnico. Después, la Pi arranca la detección al encender y revisa periódicamente las versiones estables del repositorio público eingel/cacao-detector. Sin Internet continúa con la última versión instalada. El usuario final solo abre http://:8000. En Raspberry Pi OS, copia esta carpeta deploy/ a la Pi y ejecuta una vez: El instalador crea /opt/cacao-detector/current (versión activa), /var/lib/cacao-detector/config.json (ajustes locales), cacao.service y cacao-update.timer. Necesita Internet en la primera instalación para obtener la versión estable y las dependencias. Por defecto el servo inicia en modo sim; antes de conectar el expulsor real calibra sus…

Model

cacherepair

Genglin Wang

Repair networks for CacheRepair: Learning to Repair Cross-Chunk Context in RAG for KV Cache Fusion. Each network predicts residual corrections to independently precomputed document KV caches for one frozen target LLM. The same repairer serves MuSiQue, HotpotQA, MultiHop-RAG, and TriviaQA. The Qwen repairers are Built with Qwen. The Llama repairers are Built with Llama and are named Llama-CacheRepair S/M/L. Each target/size directory contains model.safetensors and config.json. The weight file holds repair parameters and the sigmastale / sigmadelta normalization tensors in FP32. The target LLM supplies its own frozen embedding. manifest.json lists exact parameter counts and file identities.…

Model

cagliostro-v3

TobiasLogic

this is just for transfering checkpoints between instances, why are you here PLEASE STOP LIKING THIS ONE

Model · Text generation

cagliostro-v3

Bench Labs

A 146M parameter decoder-only language model pretrained from scratch on 75B tokens of open web, synthetic textbook and mathematics data. It is the third model in the cagliostro line and the first to clear an Index of 26 on the Open SLM Leaderboard metric. Zero-shot, measured with lm-evaluation-harness and the leaderboard's own ArithMark-3 script, on the exported float32 weights in this repository. The Index is the leaderboard's own formula, (N(HellaSwag,25) + N(CombinedARC,25) + N(PIQA,50) + 0.65N(ArithMark,25)) / 3.65 where N(v,c) = 100(v-c)/(100-c) and CombinedARC is the mean of ARC-Easy and ARC-Challenge. For context against other models at this scale, using the leaderboard's published…

Files (keep this directory layout -- scripts resolve imports via the parent dir): For each clip, DA3 (Depth-Anything-3) is run once over the full 77-frame GT video and once over the model's output.mp4, giving one W2C extrinsic per frame per side. Trajectories are made first-frame-relative and translation-normalized, then RotErr / TransErr / CamMC are computed (calcroterr, calctranserr, calccammc in evalcamerada3.py; math adapted from SWMbench). Per-model result = mean over clips (and per split Real/Syn). - Python env with torch, opencv-python, numpy, Pillow - depthanything3 package (Depth-Anything-3) and the checkpoint depth-anything/DA3NESTED-GIANT-LARGE-1.1 (auto-downloaded from the HF…

You should use 5123512351251235 to trigger the image generation. Download them in the Files & versions tab.

Model · Audio classification

campplus

losCrossos

This model is part of the FunASR ecosystem — one industrial-grade open-source / event · LLM-ASR. A Star really helps the project (and keeps you updated): Speaker Verification & Diarization — identify and distinguish speakers in audio. CAM++ is a speaker embedding model used for speaker verification (is this the same person?) and speaker diarization (who spoke when?). - Works with FunASR pipeline via spkmodel parameter