Skip to content
← Models
Open weightsSalesforce/10 trustVerified

CodeT5+

Salesforce's encoder-decoder code LLM family (220M to 16B) for understanding and generation, with an instruction-tuned 16B flagship.

Visit official page →

About

CodeT5+ is Salesforce Research's open family of code large language models, and one of the few serious encoder-decoder code models in a decoder-only world. It can operate in encoder-only, decoder-only, or full encoder-decoder modes, which makes it unusually flexible: code search and retrieval, summarization, translation, and generation all run on the same family. The lineup spans 220M to 16B parameters, with the larger models using a compute-efficient recipe that bootstraps from frozen CodeGen checkpoints.

The instruction-tuned InstructCodeT5+ 16B was the headline release, with the authors reporting 35% pass@1 on HumanEval, ahead of OpenAI's code-cushman-001 and state of the art among open code LLMs at the time. One caveat from the README: the instruction-tuned checkpoint used OpenAI-curated data and is licensed for research and non-commercial use only, while the base models are the open ones to build on.

Specs

Context window

Not published in surfaced materials

Pricing

Free / open weights (base models); InstructCodeT5+ 16B research/non-commercial only

Free quota

Free / open weights — self-host base models at no license cost

API access

Self-hosted via Hugging Face (Salesforce/codet5p-220m through codet5p-16b, instructcodet5p-16b)

Weights

Open weights

Released

Not disclosed

Benchmarks

  • HumanEval: InstructCodeT5+ 16B 35.0% pass@1, 54.5% pass@10, claimed to surpass code-cushman-001 (provider-reported)
  • SOTA on CodeXGLUE among open models at release (provider-reported)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

code-generationcode-retrievalcode-summarizationcode-translationcode-research

Trust score — 7/10

/10 trustVerified

Scored 7/10 (verified tier, metrics-v1). Evidence: official Salesforce Research repo (salesforce/CodeT5) with the published paper (arXiv:2305.07922); benchmark claims from the authors' own harness. Strengths: unique flexible encoder-decoder architecture, full size range, credible lab. Docked for: (1) HumanEval figures are author-run with no independent replication surfaced; (2) the instruction-tuned flagship is non-commercial, splitting the licensing story; (3) base-model license terms were not re-verified from the repo — confirm before commercial use.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
8.0/10
Manual review · sanity read + red-flag check
6.0/10
  • · Unique flexible encoder-decoder architecture across 220M-16B
  • · HumanEval figures author-run; confirm base license before commercial use

Sources