aiXcompact-4B
Context-compression model
Dynamically compresses agent history to prevent context overload; on single-GPU compute, post-training performance improves 2–3×.
MODELS
From code LLMs to specialized small models
A matrix of scenario models built for agent subtasks and for specific domain tasks.
8 YEARS · 7 GENERATIONS
The model is the product: eight years, seven generations.
RESEARCH
100+ top-venue papers, many of them pioneering work in intelligent software engineering
The aiXcoder team entered deep-learning-based program processing early, publishing some of the field's earliest academic work and contributing original results in program understanding, generation, code LLMs and agent methods.
aiXapply-4B
Code-change application model · Open source
Merges AI-generated snippets precisely back into source files: a 4B model rivaling hundred-billion-parameter models, 15× faster inference on a single consumer GPU, 30× lower task latency.
aiXcoder-7B
Code completion & generation model · Open source
SOTA code completion, published at ICSE 2025; 7B-v2 lifts API-invocation accuracy to 94.4%.
aiXverilog-4B
Embedded-development small model
Distillation and reinforcement learning bake hardware-design knowledge deep into the weights, so a 4B small model matches general large models on complex Verilog tasks — lightweight and high-performance at once.