Local Inference Lab

Independent nonprofit Open source

Intelligence,
within reach.

We develop open kernels, tools, benchmarks, and research for AI inference on hardware you control.

Explore b12x View our work
Research Kernels Tools Benchmarks
Discover
01 Open source 02 Local-first 03 Hardware-aware 04 Community-led

01 Why local

Local infrastructure should run capable AI.

Local AI improves privacy, ownership, access, and room to experiment. We build software and publish research for demanding inference workloads.

01

Performance

Hardware-conscious software extracts more inference performance from workstations and local multi-GPU systems.

02

Evidence

Reproducible benchmarks and practical research support configuration, tuning, and hardware decisions.

03

Open development

Public code and shared artifacts make results reproducible and available for independent development.

02 Areas of focus

Local AI on the most capable hardware you can own.

We focus on NVIDIA Blackwell platforms available outside traditional datacenters, with an emphasis on inference performance, efficiency, and usability.

01 / Graphics SM120

NVIDIA RTX

GeForce RTX 50 Series and RTX PRO Blackwell systems.

Consumer and professional GPUs
02 / Desktop superchip SM121

DGX Spark

GB10 Grace Blackwell systems for compact local inference.

Desktop-scale development and inference
03 / Workstation SM103

DGX Station

Workstation-class Blackwell systems for large local workloads.

Maximum local model capability

03 Flagship project

Kernels for the hardware you have.

b12x is an open SM120/SM121 CuTe DSL and Triton kernel library for local LLM inference, targeting DGX Spark and Blackwell-based RTX systems.

CuTe DSL Triton SM120 / SM121 Apache 2.0
Explore b12x on GitHub
b12x / operator map
SM120
$ pip install b12x
Plan Bind Run

04 Model deployment

Deploy a model.

Configure GLM, Qwen, DeepSeek or MiMo with the defaults from your selected LIL container. Inspect what changes and export a command for your hardware.

Open Docker configurator Read the launch guide ↗

  1. Choose a model and image

    Browse model profiles and published container builds.

  2. Set your GPU configuration

    Adjust parallelism, speculation, cache storage and memory limits.

  3. Copy a command or Compose

    Review the resolved arguments, then run the model on your own machine.

05 Join the lab

Contribute.

Run our projects on your hardware, report results, and contribute fixes through the public repositories.

GitHubCode, issues, and open development Hugging FaceModel-space work and shared artifacts X / TwitterUpdates from the lab DiscordCommunity discussion and project coordination