Salesforce's code models that actually read your variable names

What it does CodeT5 and CodeT5+ are encoder-decoder language models fine-tuned for software tasks: generating code from descriptions, completing partial functions, and summarizing code back to English. Salesforce Research released them as open models with HuggingFace checkpoints and a VS Code plugin demo.

The interesting bit The original CodeT5 (EMNLP 2021) made “identifier-aware” pre-training its signature move — treating variable and function names as meaningful tokens rather than opaque strings. CodeT5+ (2023) scales this up with larger checkpoints and reinforcement learning via the related CodeRL work. It’s a research lineage, not a product.

Key highlights

• Three concrete capabilities: text-to-code, function autocompletion, code summarization
• Pre-trained checkpoints on HuggingFace: base, large, and large-ntp-py variants
• Fine-tuned models available for downstream tasks and multilingual summarization
• BSD-3 license with an unusual ethical-use rider (no violence, environmental destruction, etc.)
• VS Code plugin demo shows real integration, though it’s labeled a research release

Caveats

• README is mostly paper links and release notes; setup instructions and benchmarks are absent
• The VS Code plugin appears to be a demo, not a maintained extension
• CodeT5+ is still an arXiv preprint as of the README’s last update

Verdict Worth exploring if you’re doing research on code LLMs or need a T5-based baseline to compare against newer decoder-only models. Skip if you want a batteries-included coding assistant — this is the raw model weights and papers.