ForgeKernel

ForgeKernel is a self-evolving GPU kernel agent that starts from a reference kernel and optimizes beyond experts.

ForgeKernel is a self-evolving GPU kernel agent that reads a reference architecture by running, timing, and profiling it. It turns the architecture into a milestone curriculum implemented by a coding agent, promoting accepted kernels to new references for continuous optimization. Findings are certified by a hardware probe and stored in a shared knowledge base. It targets NVIDIA Hopper and Blackwell GPUs, supporting languages like CuTe DSL and CUDA C++. The system features nested loops where an arch agent revises the curriculum and a coding agent implements decisions.

ForgeKernel at a glance

Best for: Software engineer
Runs on: Linux, macOS, API, CLI, self-hosted, open source
Pricing: Free
Released: October 2026
Category: AI coding tool
Website: github.com/openbmb/forgekernel

Key features

Self-evolving GPU kernel generation
Architecture-first milestone curriculum
Hardware probe certification
Shared hardware knowledge base
Nested optimization loops
Automated correctness and performance judging

Use it for

Optimizing GPU kernels past expert implementations
Generating fused attention and GEMM kernels
Benchmarking on Hopper and Blackwell architectures

Alternatives to ForgeKernel

GPUStack — Unifies AI model serving and GPU compute management across heterogeneous hardware.
ForgeCode — Leads TermBench 2.0 and runs natively right inside the terminal.
Warp — AI terminal plus cloud agents with built-in evals and benchmarks
FreeToken — Runs 290B+ MoE models on consumer gaming PCs using elastic edge resources.
Infernix — Serves massive Qwen models quickly on a single consumer GPU with advanced caching.
Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
AgentForge — It provides a flexible architecture for building specialized AI agents.
Pebrel — It brings local shells, remote hosts, and AI command-line tools into one workspace.
Enki LLM — Runs Llama-3.2 1B in pure Rust at 40 tok/s with live GPU to CPU migration.
fal — Access 1,000+ generative models and high-performance serverless GPUs.
Sandkiln — Runs untrusted code in hardware-isolated Firecracker microVMs.
supercomputer — Lower rates than AWS on-demand and Modal, billed by the second while running.

All ForgeKernel alternatives