Makora inference endpoints are built for AI agents that need to think, respond, and act in real time.

A post contributed to the AI-Driven Research for Systems (ADRS) blog series from the Berkeley Sky Computing Lab

Sequential Monte Carlo Speculative Decoding enables new highs for tokens per second per user

Writing fast FP8 GEMMs just got faster

Everyone's favourite kernel generation agent, now in your CLI!

The code was correct. The problem wasn't.

Pushing frontier model capabilities with reinforcement learning

A systematic study of reward hacking, adversarial detection, and robust evaluation for LLM-optimized GPU kernels

MakoraGenerate implements functional and fast KDA kernels with evolutionary search

Same team. Same mission. Two new letters.

Creating a representative subset of KernelBench to evaluate a long-running agent more efficiently

Announcing Makora's seed round

MakoraGenerate outperforms torch.compile when optimizing DeepSeek MOE Kernels

MakoraGenerate writes inline PTX to achieve near-optimial GEMM performance

Optimizing the kernel generation pipeline through accelerated compilation

MakoraGenerate is an LLM-powered AI agent that writes GPU kernels

Makora improves the performance of vLLM and SGLang

Achieve state-of-the-art latency on FLUX.1-schnell by leveraging multiple executor backends

Easily deploy models on Makora

Identifying the most price efficient AI inference accelerators