| language: | |
| - en | |
| license: CONFIRM_UPSTREAM_LICENSE | |
| pipeline_tag: text-generation | |
| library_name: custom | |
| tags: | |
| - code | |
| - software-engineering | |
| - on-device | |
| - gguf | |
| # AIDE Coding Model | |
| This card is intentionally incomplete until the coding-tuned checkpoint is installed and evaluated. Do not publish it with placeholder claims. | |
| ## Required Release Evidence | |
| - Exact model ID, revision, tokenizer, chat template, quantization, and runtime | |
| - Parameter count, context length, memory footprint, and measured generation speed | |
| - Coding probes for patch correctness, compile/test success, instruction following, tool-call reliability, and refusal of destructive actions | |
| - Comparison between native weights and quantized runtime output | |
| - Upstream license and all required attribution | |
| - Known limitations and supported languages | |
| ## AIDE Role | |
| The coding model is the builder lane. It may propose a unified diff, but AIDE applies no model output without path validation, diff preview, user approval, and test execution. | |