| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3-Coder-480B-A35B-Instruct |
| tags: |
| - code |
| - mlx |
| - apple-silicon |
| - local-first |
| - rmdw |
| --- |
| |
| # Super Kaiju by RMDW |
|
|
| Super Kaiju is RMDW's frontier-size private coding model, served from RMDW's own hardware as an OpenAI-compatible API. It is the big sibling to [Kaiju-Coder](https://huggingface.co/RMDWLLC/kaiju-coder-mlx-1.0): where Kaiju-Coder is the fast, local, fine-tuned builder model, Super Kaiju is a frontier-size open coder run across a two-machine Apple Silicon cluster for the work that wants the biggest model you can self-host. |
|
|
| ## What it is, honestly |
|
|
| Super Kaiju is **Qwen3-Coder-480B-A35B-Instruct** (a 480B-parameter, 35B-active mixture-of-experts coder, Apache-2.0) served at 4-bit through MLX, **pipeline-parallel across two Apple M3 Ultra Mac Studios** wired together over Thunderbolt with RDMA. It is wrapped in RMDW's Kaiju harness and system identity and exposed as a drop-in OpenAI-compatible endpoint. |
|
|
| It is not a fine-tune and it is not claimed to beat the frontier labs. The honest pitch: |
|
|
| - **Frontier-size, fully private.** A 480B coder running on hardware you (RMDW) own, no tokens leaving the building. |
| - **Flat-rate, not metered.** One monthly price, send the requests your work needs. |
| - **Capability over speed.** Because it is a frontier-size model on a home cluster, it trades throughput for size: expect roughly **13 tokens/second**, slower than the entry Kaiju-Coder. You run Super Kaiju when you want the larger model, not when you want the fastest one. |
|
|
| ## Model details |
|
|
| - **Base model:** [Qwen/Qwen3-Coder-480B-A35B-Instruct](https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct) (Apache-2.0) |
| - **Served quant:** [mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit](https://huggingface.co/mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit) (~270 GB) |
| - **Serving:** MLX pipeline-parallel across 2× Apple M3 Ultra (256 GB each), JACCL/RDMA over Thunderbolt, custom main-thread inference server |
| - **Context:** 256K (extendable) |
| - **Interface:** OpenAI-compatible `POST /v1/chat/completions`, streaming and non-streaming |
| - **License:** Apache-2.0, fine-tuned-from / served-from Qwen under Apache 2.0. No Qwen or Alibaba endorsement implied. |
|
|
| ## Intended use |
|
|
| Building complete websites and apps, wiring Stripe and auth, writing scripts and automations, drafting proposals and invoices, and the rest of the real work that ships a one-person business, on a private endpoint with no per-token meter. |
|
|
| ## How to use |
|
|
| ```bash |
| curl https://api.rmdw.ai/v1/chat/completions \ |
| -H "Authorization: Bearer rmdw-YOUR_KEY" \ |
| -H "Content-Type: application/json" \ |
| -d '{"model":"super-kaiju","messages":[{"role":"user","content":"Write a Python function that returns the nth Fibonacci number."}]}' |
| ``` |
|
|
| Built and self-hosted by Richard Echols / RMDW LLC, Atlanta. |
|
|
