30 entries · Checked 2026-10-10
Companies
Labs, hardware makers and toolmakers betting on AI that runs on your own machine.
Hugging Face
Hosts the models, datasets and quantizations the whole scene depends on.
huggingface.co labFRMistral AI
Paris lab that releases most of its models under Apache 2.0, from Ministral 3B to Mistral Large 3.
mistral.ai labFRKyutai
Non-profit lab in Paris. Open speech models like Moshi that talk back in real time.
kyutai.org labDEBlack Forest Labs
The FLUX image models, the default pick for local image generation.
huggingface.co/black-forest-labs labUS · GBGoogle Gemma
Gemma 4 covers phones (E2B) up to a single 24 GB card (31B), all Apache 2.0.
huggingface.co/google labCNQwen (Alibaba)
Every size from phone to server, mostly Apache 2.0, and usually the first to get quantized.
huggingface.co/Qwen labCNDeepSeek
MIT-licensed MoE models that pulled open weights close to the frontier.
huggingface.co/deepseek-ai labCNMoonshot AI
The Kimi line. Trillion-parameter MoE with open weights.
huggingface.co/moonshotai labCNZ.ai
GLM models, strong at coding and agent work.
huggingface.co/zai-org labUSNous Research
Hermes fine-tunes and open research on distributed training.
nousresearch.com runtimeUSOllama
One command to pull and run a model. The easiest way in on any OS.
ollama.com runtimeUSLM Studio
Desktop app to find, download and chat with local models, with an OpenAI-compatible server built in.
lmstudio.ai toolingUSComfy Org
The company behind ComfyUI. Day-one workflows for new image and video models, all running locally.
comfy.org toolingUSUnsloth
Fine-tuning that fits on one consumer GPU, plus carefully tested GGUF quants.
unsloth.ai hardwareUSNVIDIA
CUDA is still the first target of every inference engine.
nvidia.com hardwareUSAMD
ROCm keeps improving, and Ryzen AI Max puts 128 GB of unified memory in a small desktop.
amd.com hardwareUSIntel
Arc cards offer a lot of VRAM per euro and an open software stack.
intel.com hardwareUSApple MLX
Unified memory on M-series Macs plus MLX make large models practical on a desk.
github.com/ml-explore/mlx hardwareUSFramework
Framework Desktop, a small repairable PC built around Ryzen AI Max with up to 128 GB.
frame.work hardwareUStiny corp
tinygrad and the tinybox, multi-GPU machines that ship ready for local training and inference.
tinygrad.org labUSEleutherAI
Non-profit lab behind the Pythia and GPT-NeoX open models and datasets, and a mainstay of open-model training research.
eleuther.ai labUSAllen Institute for AI (Ai2)
Releases fully open models such as OLMo together with training data and code, so they can be run and reproduced locally.
allenai.org labAETechnology Innovation Institute (TII)
Abu Dhabi research centre behind the Falcon open-weight model family.
tii.ae hardwareUSTenstorrent
Builds RISC-V-based AI accelerators and dev kits, an inference path that is neither CUDA nor ROCm.
tenstorrent.com hardwareUSQualcomm
Snapdragon and Hexagon NPUs put local LLM inference into laptops, phones and small boards.
qualcomm.com runtimeUSJan
Open-source desktop app that runs GGUF models locally behind an OpenAI-compatible server.
jan.ai runtimeITLocalAI
Container-friendly engine that serves LLMs, vision and speech models locally with an OpenAI-style API.
localai.io runtimeUSSGLang
High-performance serving framework for LLMs and multimodal models, widely used for self-hosted inference.
github.com/sgl-project/sglang toolingUSGPT4All
Desktop chatbot and local server that runs quantised models fully offline on CPU or GPU.
gpt4all.io hubUSCivitai
Large sharing hub for the Stable Diffusion and Flux checkpoints, LoRAs and VAEs that local image users download.
civitai.com