Kapitsa

Kapitsa is an inference engine optimized for Apple Silicon that runs Qwen and Nemotron models locally on Mac hardware.

Kapitsa is a local inference engine designed for Apple Silicon Macs with macOS 14 Sonoma or later. It runs specific Qwen and Nemotron models, including dense and mixture-of-experts architectures. Users select models and precisions via the command-line interface, with options ranging from 4-bit up to unquantized formats. The engine processes text inputs and, for most supported models, image inputs. It provides hardware-optimized performance benchmarks across various MacBook Pro configurations.

Kapitsa at a glance

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

Key features

Run Qwen and Nemotron models locally
Support Apple Silicon Macs with macOS 14
Select models and precisions via CLI
Process text and image inputs
Execute benchmark tests and submit results
Import custom GGUF model formats

Use it for

Running local AI models on Mac hardware
Benchmarking local LLM inference speeds
Executing coding and text generation tasks
Processing multimodal inputs with local models

Alternatives to Kapitsa

Pulsar — It is the fastest inference engine for Qwen models on Apple Silicon.
oMLX — It uses paged SSD KV caching to reduce agent TTFT to under 5 seconds.
MTPLX — It runs local models twice as fast via native MTP speculative decoding.
Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
PocketWebTools for Mac — Runs local AI models up to 35B parameters fully offline on Apple Silicon Macs.
Strata Mlx — Runs 125B MoE models on Apple Silicon by streaming experts directly from disk.
Union.ai — Runs machine learning tasks natively on Apple Silicon GPUs using Metal.
Kaku — It works immediately out of the box with zero configuration required.
Rapid-MLX — Run 250 MLX-native models locally on Apple Silicon with high inference speeds.
Reflex — It is optimized for cold-start latency and single-tenant decisions.
fal — Access 1,000+ generative models and high-performance serverless GPUs.
Infernix — Serves massive Qwen models quickly on a single consumer GPU with advanced caching.

All Kapitsa alternatives