Post
95
Wisp Coder is public, and you can now try it in your browser. 🪶
I trained this small code model from scratch to explore fill-in-the-middle and native multi-token prediction on Apple Silicon: 108.2M parameters including its MTP module, with a 100.7M Llama trunk.
The new demo lets you edit the code around a cursor and generate a suggestion. It runs the standard trunk on free CPU hardware; it does not use MTP acceleration or execute generated code.
One lesson from this work: draft acceptance alone does not establish better code or lower latency. The published evaluations include null and adverse results, and the registered short-hole study did not establish superiority over the compared FIM models. Wisp is a small research model with limited completion quality.
Try the demo: philipjohnbasile/wisp-coder-demo
Model, methods, and results: philipjohnbasile/wisp-coder-110m
More of my Apple Silicon work: philipjohnbasile/start-here-selected-work-6aa309b179b2a196db52a12b
If you work on FIM, MLX, or local autocomplete, I'd welcome feedback and reproducible examples.
I trained this small code model from scratch to explore fill-in-the-middle and native multi-token prediction on Apple Silicon: 108.2M parameters including its MTP module, with a 100.7M Llama trunk.
The new demo lets you edit the code around a cursor and generate a suggestion. It runs the standard trunk on free CPU hardware; it does not use MTP acceleration or execute generated code.
One lesson from this work: draft acceptance alone does not establish better code or lower latency. The published evaluations include null and adverse results, and the registered short-hole study did not establish superiority over the compared FIM models. Wisp is a small research model with limited completion quality.
Try the demo: philipjohnbasile/wisp-coder-demo
Model, methods, and results: philipjohnbasile/wisp-coder-110m
More of my Apple Silicon work: philipjohnbasile/start-here-selected-work-6aa309b179b2a196db52a12b
If you work on FIM, MLX, or local autocomplete, I'd welcome feedback and reproducible examples.