Drex Max is here: the decision model that sees.
Nace.AI
Talk to salesSign inGet started
Kapitsa

The fastest way to run Qwen and Nemotron on a Mac.

A local inference engine built for two model families and one chip.

brew install nace-ai/tap/kapitsa
View on GitHubApple silicon · macOS 15 or later

Same answer, real time

3.3 s
  1. Kapitsa154 tok/s
  2. oMLX + DFlash2127 tok/s
  3. Ollama89 tok/s
  4. LM Studio86 tok/s
  5. MTPLX80 tok/s
  6. llama.cpp69 tok/s
Qwen3.6 35B-A3B at 4-bit, writing a coding answer on a MacBook Pro with M4 Pro and 48 GB.

Faster on every model it runs.

Tokens per second writing a coding answer. Median of five runs, every engine at its latest release.

Qwen3.6 35B-A3B · 4-bit · MacBook Pro, M4 Pro, 48 GB

  1. Kapitsa154 tok/s
  2. oMLX + DFlash2127 tok/s
  3. mlx-serve + MTP124 tok/s
  4. mlx-serve + DFlash2104 tok/s
  5. Ollama89 tok/s
  6. LM Studio86 tok/s
  7. MTPLX80 tok/s
  8. llama.cpp + MTP69 tok/s
Follow-up at 16K
0.09 s
Best other: MTPLX, 0.15 s
Peak memory
21.1 GB
Best other: mlx-serve, 22.5 GB
How it was measuredMeasure your Mac with kapitsa bench

Two families. Five models.

kapitsa doctor shows which models fit your Mac.

Qwen

  • Qwen3.5 4B

    Dense · For 16–24 GB Macs

    3.3 GB

    --model 4b

  • Qwen3.6 35B-A3B

    MoE, 3B active · Quickest to reply

    22.1 GB

    --model 35b

  • Qwen3.8 27B

    Dense · Strongest coder

    20.2 GB

    --model 27b

Nemotron

  • Nemotron 3.5 Bolt 74B-A9B

    Hybrid Mamba-2 MoE, 9B active · Built for agents

    25.1 GB

    --model bolt

  • Nemotron 3.5 Super-VL 120B-A12B

    Hybrid Mamba-2 MoE, 12B active · For 96 GB+ Macs

    76.4 GB

    --model super

Bring your coding agent.

One command downloads the model, starts it and opens your agent on it. Any OpenAI or Anthropic client works too.

An oil painting of a person working at a vintage computer on a hillside below a mountain
kapitsa claude
Base URL
http://127.0.0.1:8000
POST /v1/chat/completions
OpenAI
POST /v1/responses
OpenAI
POST /v1/messages
Anthropic

Built around the model.

Hand-written Metal kernels for each weight format. No MLX or llama.cpp underneath.

  • Reads on two chips

    Prompts are read on the GPU, and on dense models and Nemotron the Neural Engine reads part of each layer beside it.

  • Writes several tokens at once

    A draft proposes a block of tokens. The model checks them all in one pass.

  • Follow-ups read only what’s new

    Each conversation keeps its cache, so a follow-up at 16K tokens starts in 0.09 s on the 35B.

Architecture walkthrough
An oil painting of a sunset over a still lake lined with trees

Install Kapitsa.

Free and open source under Apache 2.0.

brew install nace-ai/tap/kapitsa
kapitsa claude
View on GitHubApple silicon · macOS 15 or later