JustVugg / colibri

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

mw00 / project-maya

Project Maya - run GLM-5.3-Flash (321B MoE) and DeepSeek-V4.1-Flash-1017 (748B MoE) privately on your own GPU(s), from one card to 16, NVIDIA or AMD, Linux or Windows. Engine, server, dashboard, OpenAI- and Anthropic-compatible API, quant tools.

bojieli / ai-infra-book

《深入理解 AI Infra：量化分析与系统设计》（李博杰 著）开源书稿：从硬件约束和模型架构出发，量化推导 LLM 推理与训练系统设计。含全书正文、PDF、配套计算工具与实验

VecSzn / PolyStrata

Run large AI models on your own PC: Qwen3.8-Flash-Next on Strata's engine, and GLM 5.3 Flash, DeepSeek V4 Flash, LFM2.5, Qwen3.6, Ornith 1.5, Gemma 4, Tiel Coder and MiniCPM-V on a second engine written for PolyStrata. One-click install for Windows and Linux, OpenAI/Anthropic API on localhost.

Indras-Mirror / Xolotl

Xolotl (formerly Strata-DSV4.1): DeepSeek-V4.1-Flash, all 552B, on one gaming PC (RTX 4090 + 90 GB RAM) at ~27 tok/s, plus GLM-5.3-Flash, Qwen3.8-Flash-Next and Qwen3.8-27B. A heavily extended fork of Strata: MoE experts tiered across VRAM, pinned RAM and NVMe; OpenAI + Anthropic API.

FareedKhan-dev / kimi-k3-in-c

A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.

peonist-ai / halogen-flash-server

The fastest way to run Qwen3.8-Flash-Next on Strix Halo (gfx1151)

meta-pytorch / dist_moe

Dist-MoE implements Speed-of-Light distributed Mixture-of-Experts forward and backward kernels. They are highly optimized and fused comm-compute kernels for training and inference on SM100+ GPUs

deepspeedai / DeepSpeed

DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.

prabanta-dev / janas

37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.

ARahim3 / kaggle-tpu-lab

Frontier-class open models on a free Kaggle TPU v5e-8: GLM-5.3-Flash 320B MoE (~165 tok/s, our own JAX engine) and Qwen3.8-27B bf16 (~130 tok/s), 262k context, prefix caching. Works with Claude Code, Codex, opencode and pi.

Helldez / BigMoeOnEdge

Run MoE models bigger than your RAM. Frontier-size MoE on a phone, laptop, edge. Lossless, on stock llama.cpp. An adaptive streaming optimizer. Any hardware, any OS, any model, any quantization, behind a local OpenAI-compatible server.

midagedev / bloomery

Hybrid GPU + CPU inference for mixture-of-experts models, written in Rust down to the CUDA kernels. A drop-in llama-server.

NVIDIA / cudnn-frontend

cuDNN Frontend is NVIDIA's modern, open-source entry point to the cuDNN library and a growing collection of high-performance open-source kernels.

cursor / mixture-of-kittens

Mixture-of-experts (MoE) training megakernel for NVL72s

algorithmicsuperintelligence / optillm

Optimizing inference proxy for LLMs

Tylogi / TyloQuant

Get more intelligence from every bit. Better quantization formats and smarter calibration let larger, stronger models run smoothly on the hardware you already own.

carloslfu / slotstream

Run a 105 GB AI model on a Mac that can't hold it. Slotstream streams Qwen3.8-Flash-Next (125B mixture of experts) from your SSD and caches the busiest experts in memory, so it runs on Macs with 16 to 64 GB. One native Swift binary on MLX and Metal, no Python, offline. Works with Claude Code, Codex and Ollama or OpenAI clients.

JustVugg / lumabri

Run large AI models across your computers and servers. Powered by Colibri. Pure C.

1115714829 / llama.cpp-v100

llama.cpp optimized for NVIDIA Tesla V100 (SM70): Qwen3.8-Flash-Next with MTP (prefill 3402 tok/s at 200K context on 6x V100-16GB; also runs on 2x V100-16GB) and Qwen3.8-27B with DFlash2 (decode up to 220 tok/s); 2-6 GPUs, 262K-512K context, multimodal, concurrent serving, built-in web UI.

gpillon / ignis

OpenAI-compatible inference server for Qwen3.8 on one RTX 5090 / RTX PRO 6000: the 27B at ~170 tok/s (peaks ~370), Flash-Next (125B MoE) at ~100 tok/s, built for local coding agents. Rust + CUDA, Windows and Linux.

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.

DingoOz / llm-visuals

Live truecolor terminal dashboard for a locally running LLM: tok/s, TTFT, GPUs, context, MTP acceptance, layer and real expert routing (llama.cpp)

surya-koritala / condor

Qwen3.8-Flash-Next (180B MoE) on one 12-16 GB NVIDIA GPU, Linux or Windows. Own engine, own compression.

microsoft / Tutel

Tutel MoE: Optimized Mixture-of-Experts Library, Support GLM/Kimi/Gpt/DeepSeek/Qwen using FP8/NVFP4/MXFP4

tuff-app / TUFF

The local AI app built for the Mac. Run LLMs like Qwen, Gemma and GPT-OSS on Apple Silicon, including models far bigger than your Mac's memory, by streaming MoE experts from the SSD.

PedroMarinhoDev / ComfyUI-HunyuanImage3

Run Tencent's HunyuanImage-3.0 (80B MoE) natively in ComfyUI on a 24 GB GPU: base, Instruct and Instruct-Distil, prompt rewriting, image editing, Spectrum speed-up.

SMTorg / smt

SMT: The Surrogate Modeling Toolbox

brontoguana / krasis

Krasis is a Hybrid LLM runtime which focuses on efficient running of larger models on consumer grade VRAM limited hardware

NeelM0906 / Mference

Swift + Metal MoE inference for Apple Silicon: Qwen 3.6 35B at 23.5–29.3 tok/s decode with 2.20× faster long-prompt prefill on a 24 GB M5; Gemma 4 26B in ~2 GB, DeepSeek-V4-Flash 284B, Inkling-Small 276B, native Mac app, CLI, and OpenAI-compatible server.

1picassoai / coachwhip

The 80B Qwen3-Coder-Next on a 16 GB Mac, about 13 tok/s, fully offline. A mixture-of-experts inference engine for Apple silicon.

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).

TimRots / kimi3

Independent from-scratch implementation of the Kimi K3 architecture (arXiv:2607.24653v1): KDA, NoPE, latent-space MoE, and the systems co-designs. Table 1 reproduced to 0.09%.

tardellirs / gemma3moe-builder

Sparse-upcycle a dense Gemma-3-270m into a Mixture-of-Experts (gemma3moe) and serve it on llama.cpp (CPU). Upcycle, train, inspect routing, GGUF.

shiloz-stack / blackwell-moe-kernel-scheduler

Workload-aware persistent expert-tile scheduling for irregular MoE inference kernels on NVIDIA Blackwell GPUs

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.

shyringo / deepseek-v4-flash-0731-in-c

Run the native 284B-A13B DeepSeek-V4-Flash-0731 LLM locally on one laptop CPU: pure C, 8 GB RAM minimum, no GPU, best TPOT 0.892 s/token, resident OpenAI-compatible API with function tools. | 在笔记本单颗 CPU 上本地运行原生 284B-A13B DeepSeek-V4-Flash-0731 大模型：纯 C，最低 8 GB 内存，无需 GPU，最优 TPOT 0.892 秒/token，支持模型常驻的 OpenAI 兼容接口与函数工具。

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.