**openai/gpt-oss-120b** — MLX (MXFP4), ~66 tokens/sec @ Hugging Face: `lmstudio-community/gpt-oss-120b-MLX-8bit`
**google/gemma-3-27b** — MLX (4-bit), ~27 tokens/sec @ Hugging Face: `mlx-community/gemma-3-27b-it-qat-4bit`
**qwen/qwen3-coder-30b** — MLX (8-bit), ~78 tokens/sec @ Hugging Face: `Qwen/Qwen3-Coder-30B-A3B-Instruct`
Will reply back and add Meta Llama performance shortly. Models
gpt-oss-120b, Meta Llama 3.2, or Gemma
(just depends on what I’m doing)
Hardware
- Apple M4 Max (128 GB RAM)
paired with a GPD Win 4 running Ubuntu 24.04 over USB-C networking
Software
- Claude Code
- RA.Aid
- llama.cpp
For CUDA computing, I use an older NVIDIA RTX 2080 in an old System76 workstation.
Process
I create a good INSTRUCTIONS.md for Claude/Raid that specifies a task & production process with a task list it maintains. I use Claude Agents with an Agent Organizer that helps determine which agents to use. It creates the architecture, prd and security design, writes the code, and then lints, tests and does a code review. KickBack Rewards Systems
http://www.kickbacksystems.com
http://careers.kickbacksystems.com
KickBack Rewards Systems (KRS) is a bootstrapped company that specializes in customer specific marketing and payments solutions for over 1000 US clients. Our software teams have been remote-first since the early 2000s and we will always be remote.
At first I thought this might be a Nano SaaS, but without a clear definition, I’m now guessing that would be something like one tenth of the numbers you mentioned. So “micro” seems right.
If the term or threshold does not exist, let’s define it as such.