THUSI-Lab / GameScaling

Scaling in Games: Continued Pre-Training for Embodied Agents in Diverse Virtual Worlds. Weights on Hugging Face.

Edge0-AI / Edge0

Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported

StartLuxLabs / StartLux-Decision

Open decision models from 0.8B to 27B: a calibrated probability for every option in one pass, and a short thought only when unsure. Engine for vLLM, llama.cpp and MLX.

InlitX / GymMane

Offline Android gym log for tracking sets, weights, routines and progress, with no account, ads, or internet permission.

skilltraining-project / skill-training

Skill training with a corruption and reconstruction loop: train an agent's skill library, not its weights.

hi3d-official / Hi3Dv1.0

Official Hi3D v1.0: high-resolution 3D geometry refinement built on Sparc3D. Inference code, model weights, and documentation.

raullenchai / Rapid-MLX

Rapid-MLX is an open-source (Apache 2.0) OpenAI- and Anthropic-compatible LLM inference server for Apple Silicon, built on MLX, focused on reliable tool calling for coding agents. Up to 4× faster than Apple's MLX (mlx-lm) on the same weights.

leochlon / bitsy

Prompt or weights: gate, bakeoff, diagnose, guard.

banadmiralpurge / lora-merge

Published: 2026-10-18

black-forest-labs / flux-action

FLUX 3 Action: open weights 7B world action model from Black Forest Labs. Train, fine-tune and run action prediction for robots (DROID, SO-101), simulators and games.

xiaoranlingfeng / ppt-master-zit

ppt-master (v6.7.0) fork: local ComfyUI image backend (Qwen-Image-2.1 default, Z-Image Turbo selectable) + fully editable PPTX workflow. No API key, no per-image cost; Qwen weights are non-commercial — see README.

Wassimyounes01 / qwen38-uncensored

Local Qwen 3.8 27B uncensored Q4_K_M + harvested SYSTEM pack. Official 3.8 weights, not a 3.6 retitle.

speridlabs / iris-3b

Iris-3B: pixel-space generation & general vision learner. A 3B diffusion transformer that generates every pixel directly (no VAE); the same prior, fine-tuned, does monocular depth estimation and 4x image restoration. Paper, weights, training code and demo.

huggingface / pytorch-image-models

The largest collection of PyTorch image encoders / backbones. Including train, eval, inference, export scripts, and pretrained weights -- ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (ViT), MobileNetV4, MobileNet-V3 & V2, RegNet, DPN, CSPNet, Swin Transformer, MaxViT, CoAtNet, ConvNeXt, and more

qubvel-org / segmentation_models.pytorch

Semantic segmentation models with 500+ pretrained convolutional and transformer-based backbones.

danveloper / flash-moe

Runs a large language model on a small laptop by streaming weights from SSD, written in Objective-C.

ultralytics / assets

Shared Ultralytics logos, media, sample images, pretrained model weights, dataset artifacts, and release assets.

Amey-Thakur / CLAUDE-CERTIFICATIONS

Study guide for all four Anthropic Claude certifications: Associate Foundations (CCAO-F), Developer Foundations (CCDV-F), Architect Foundations (CCAR-F), Architect Professional (CCAR-P). Official exam guides, blueprints with domain weights, cheat sheets, 320 practice questions, three mock exams each, 110 flashcards, and a printable companion.

igrbible / Ruach_Studio

Unbelievable Ruach Studio Sound Heresy. Borderless Open Weights SUNO Killer based on YuE2 SOTA Model. Creator, Writer, Refiner, Artist, Librarian, Trainer — Six as One. Everything you missed in SUNO and such. Turn the songs of your Heart into thundering energy — clear and unmuffled. Thus we call it OSEM: Open Source, Engaged & Musical.

DominikBucko / qwen38-flash-next-2x3090

Qwen3.8-Flash-Next on 2× RTX 3090: with 128 GB RAM up to 4,191 tok/s prefill · 111.5 tok/s decode (131K prompt), full 256K window at 2,865 tok/s; with 64 GB RAM 3,410 tok/s prefill · 84 tok/s decode. New: opt-in uncensored mode (runtime abliteration, no new weights).

iamwavecut / MLX-DLSS

NVIDIA DLSS 5 neural rendering and DLSS frame generation on Apple Silicon (MLX/Metal, Core ML) and PyTorch; weights extracted from your own DLSS libraries

msuiche / weightless

Abliteration without the weights — projective refusal steering as a GLP vector + fail-closed boot hotfix.

professorpalmer / bonsai-ada-surgery

Bonsai 2 27B: 262k context (trained max) on 12 GB, 54 to 97 tok/s decode, 2x prefill, q8_0 KV recipe, grammar tool calls 9/9. Same ternary weights. Windows bundle + receipts.

Sam-r-passmore / energy-icons

Open-source icon library for the energy transition: 1,200+ icons in two optical sizes and two weights. MIT.

wandb / wandb

The AI developer platform. Use Weights & Biases to train and fine-tune models, and manage models from experimentation to production.

sqliteai / warp

Run the full 2.78-trillion-parameter Kimi K3 model, DeepSeek V4.1 Flash or GLM-5.3-Flash beyond available RAM by streaming activated weights directly from NVMe. A dependency-free, embeddable C inference engine.

magenta / magenta-realtime

Magenta RealTime 2: An Open-Weights Live Music Model

leonickson1 / Swiftlet

Swiftlet is a Swift and Metal runtime that runs large Qwen Mixture-of-Experts models locally on Apple devices by streaming expert weights from storage, enabling 35B and 80B models to run with low RAM, including on iPhone.

google-gemma / awesome-gemma

😎 Awesome list about Gemma, Google DeepMind's family of lightweights, state-of-the-art open models.

programasweights / claudish

Translate between English and Claudish with tiny ProgramAsWeights functions.

allebee / jevk5

JevK5: open-weight alternative to TypeSafe Jev. Typed decisions with probabilities in one forward pass; Apache-2.0 weights and code.

sybil-solutions / deepseek-v4.1-flash-4x-rtx-pro-6000

DeepSeek V4.1 Flash on 4x RTX PRO 6000 Blackwell: native weights, DSpark, vision, NVMe or locked RAM Engram offload.

bertholomus / deepseek-v4.1-tensorfold-tp2-2xgb10

BertholomusAI recipe: DeepSeek-V4.1-Flash on 2x NVIDIA DGX Spark (GB10), our own TensorFold TP2 engine. Weights: Mia-AiLab EXL3 2.9bpw (MIT).

ModelTC / LightVAE

LightVAE: Fast video VAE decoders for Wan2.1, Wan2.2, and MiniMax-H3, plus a compressed H3 reference encoder. Pretrained weights and ComfyUI workflows.

adobe-fonts / source-serif

Typeface for setting text in many sizes, weights, and languages. Designed to complement Source Sans.

ikermoel / open-alternative-jev

Open-source alternative to TypeSafe's Jev: a System One style model layer that gives typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with honest benchmarks

sudoingX / dgx-spark-deepseek-v41

DeepSeek-V4.1-Flash on two NVIDIA DGX Sparks: EXL3 weights from the official checkpoint, upstream vLLM with our patches and decode kernels

Greninja9257 / LabLLM

A native macOS lab for teaching tiny language models to think — build the architecture, train the weights, and watch a small LLM emerge from scratch, locally on Apple Silicon with custom data, tokenizers, checkpoints, and MLX acceleration.

sepiablue-ai / ComfyUI-MiniMax-H3-TaoMate-C3

TaoMate-H3 3-step acceleration package for ComfyUI on an RTX 4070 12GB: 4-bit (W4A4) FFN weights, 10% sparse attention and extra INT8 weight caching, with Windows setup and sample workflows.

wandb / weave

Weave is a toolkit for developing AI-powered applications, built by Weights & Biases.

MoonshotAI / checkpoint-engine

Checkpoint-engine is a simple middleware to update model weights in LLM inference engines

hmorimitsu / ptlflow

PyTorch Lightning Optical Flow models, scripts, and pretrained weights.

sepiablue-ai / ComfyUI-MiniMax-H3-W4A4-VSA

MiniMax H3 Ref2VA acceleration for ComfyUI on an RTX 4070 12GB: 4-bit FC1 weights + Streaming VSA sparse attention. 1024x1792 video with audio in about 3 min 17 s, about 3.4x faster than dense attention.

Sparticle62ops / pssa

PSSA: a non-transformer language model written from scratch in Rust. Recurrent state-space layer, episodic memory, self-modifying weights, CUDA training, and a full terminal dashboard.

atbender / deepseek-v41-flash-mac-mini

Run original DeepSeek V4.1 Flash FP4/FP8 weights on a 16 GB M1 Mac mini with bounded SSD streaming in MLX. Recipe, benchmarks, and demo.

sahibzada-allahyar / gliner2-ultrafast

Real browser automation with local GLiNER2 inference and open weights. Work in progress. Contributions welcome.

nanguoyu / minirun-app

Run Kimi K3 (2.8 T parameters,1.56 TB) on iPhone 16 Pro? Weights stream from your SSD through a memory budget you set. Free, local, on Mac and iPhone.

ortusaitech / rightwayup

Full-circle image roll estimation with calibrated abstention. Open weights, Apache-2.0. By ORTUS AI, the team behind CHEQIT.

fulvian / engraft-ngram

ENGRAFT: write facts into an LLM's n-gram (Engram) memory table as a removable overlay: 100 facts in a 9 MB file, no weights touched, verified on llama.cpp GGUF (Qwen3.8-Flash-Next).

Vxtzq / CrowdGPT

CrowdGPT is an opensource AI that aims to be trained without any datacenters

ownweights / argus

Open-source QA agent swarms. Argus tests web apps like a human tester would but even more accurate and faster. Visual navigation, real workflows, evidence-backed reports.

lukeckprobierts / Hemmingway-1

Hemmingway-1: a 27B model that writes the way a person writes. Open weights, free for non-commercial use (CC BY-NC 4.0); commercial use by agreement.

IamBusy / OpenJev-Vision

Open visual probability research: encode an image once, answer multiple structured questions. Public data, trained weights, reproducible baselines and failure cases.

franciscocarloserra / ttslibre

Small, fast, free: an open recipe for a sub-100M TTS. Code, data, weights. The PoC trained to a working model in a few hours on one RTX 3090, and inspired a wave of open-weights TTS releases.

sauravsingla / MemVanta

Low-memory C++20 LLM inference runtime for quantized GGUF models on CPU — mmap-backed weights, paged KV cache, Q4/Q8 kernels, and reproducible llama.cpp benchmarks.

programasweights / pii

Detect and type PII locally with one ProgramAsWeights neural program.

WestWaters / pollard-weights

A tiered-memory system design for workloads that don't fit in RAM: measure the working set, pin the hot tier, stream the cold tier from flash. Ships the residency calculator, measurement harnesses, and the build recipes behind it. Predictions validated against public benchmarks.

vcruz305 / vllm-exl3

Serve EXL3 (ExLlamaV3 trellis) quantized models on vLLM fork runtimes — any architecture, mixed per-layer bitrates, composable with source-format non-routed weights

rayl15 / Godwit

Run GPT-OSS-120B on a 16 GB Mac. Streams mixture-of-experts weights from SSD so only the 2 GiB trunk stays resident. Swift + Metal, no dependencies.

Uzbekunknown / dlss-nr-on-intel

DLSS 5 Neural Rendering on an Intel Xe2 iGPU: an independent reimplementation of the inference pass, running in a real game through a Vulkan layer. Code only. No NVIDIA binaries, no weights.