Software

Software frameworks

    • Cogent Core: A cross-platform framework for building powerful, fast, elegant 2D and 3D apps

    • Emergent: Neural network simulation infrastructure, built on top of Cogent Core.

    • Axon: Spiking neural network simulation algorithms, used for current models in compcogneuro.org

See zotero for usage and configuration of zotero for paper and reference management.

AI on MacBook Pro

I happen to have an M3 Max chip with 128 GB of ram (always max out your system! sure paid off now given RAM prices!), and can therefore run a decent LLM locally. After sorting through the usual reddit threads etc, here’s what I ended up with (Oct, 2026):

    • https://github.com/ddalcu/mlx-serve – because they seem to have the sweet-spot for the best quantized version of the Qwen3.8-Flash-Next model. My trajectory went through: https://lmstudio.ai/ Bionic app (which seemed very slow), then https://omlx.ai/ https://github.com/jundot/omlx which only had a 95Gb version of the model that was too big and caused conflicts with running other apps (install tips: key is to follow the homebrew instructions – even though I use xcode command line tools, had to reinstall that – picky. Even so, could not use the --with-custom-kernel – build failed – no idea what that actually does anyway).

    • http://localhost:11234/ (once server is up) has good settings and model download GUI.

    • good models: ddalcu/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw fits in 70Gb, whereas Qwen3.8-Flash-Next-oQ4e-mtp takes up 96Gb which was too much, and Qwen3.8-27B-oQ4e-mtp which fits in 16Gb, but is less smart and runs more slowly. https://huggingface.co/ddalcu/DeepSeek-V4-Flash-0731-iQ-MLX-3.3bpw does better on the https://arena.ai/leaderboard/agent/code benchmarks, but it takes 105Gb and is therefore not usable. It is also slower.

    • Using OpenCode which seems good, and integrates with everything: https://opencode.ai/ https://github.com/anomalyco/opencode – they should have a new v2 version soon.

    • The mlx-serve launch opencode or omlx launch opencode all seemed to fail (e.g., OpenCode is not installed. Install: curl -fsSL https://opencode.ai/install | bash even after I had done that already), so I followed the google AI instructions and just hand-edited the .json file below, which worked.

    • Some other people recommend https://github.com/beamivalice/sushi (which is a fork of mlx-serve) instead of oMLX – allows the flash model to be quantized down further for smaller RAM systems. e.g., https://huggingface.co/beamster/Qwen3.8-Flash-Next-Sushi-4bpw is 64Gb.

    • The Flash-Next seems both fast and smart enough to actually be quite usable! I will continue to test it – not as fast and smart as Claude Opus 5.5 but still it seems reasonably capable.

OpenCode config file for mlx-server

~/.config/opencode/ > cat opencode.json

{
  "provider": {
    "mlx-server": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "mlx-server",
      "options": {
        "baseURL": "http://localhost:11234/v1"
      },
      "models": {
        "Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw": {
          "name": "Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw",
          "limit": {
            "context": 262144,
            "output": 65536
          }
        }
      }
    }
  },
  "model": "mlx-server/Qwen3.8-Flash-Next-MLX-Serve-iQ-MLX-4.7bpw"
}