CodeT5-base recorded 22,172 monthly downloads on Hugging Face as of 2026, according to platform data. Salesforce Research built this encoder-decoder transformer family across eight model variants ranging from 60 million to 16 billion parameters. The instruction-tuned CodeT5+ 16B scored 35.0% pass@1 on the HumanEval benchmark in zero-shot settings, beating OpenAI’s code-cushman-001 at 33.5%. This article covers the full scope of CodeT5 statistics in 2026, including code generation accuracy, developer adoption patterns, training data composition, and benchmark results.
CodeT5 Statistics – TL;DR
CodeT5-base generates 22,172 monthly downloads on Hugging Face, with 761+ related models deployed across the platform as of 2026.
InstructCodeT5+ 16B reached 35.0% pass@1 and 54.5% pass@10 on HumanEval benchmarks in zero-shot code generation.
CodeT5+ trained on 51.5 billion tokens from GitHub repositories, a 50x increase over the original CodeSearchNet corpus of 8.35 million instances.
The CodeT5+ 770M model matched models 8 to 80 times its size, hitting 15.5% pass@1 on par with PaLM 62B and GPT-NeoX 20B.
Salesforce archived the official CodeT5 GitHub repository in May 2025. Model weights remain available on Hugging Face under Apache 2.0 licensing, and community forks are actively maintained.
How Many Downloads Does CodeT5 Get on Hugging Face?
CodeT5-base pulled 22,172 monthly downloads on Hugging Face as of 2026. The model accumulated 132 community likes and powers 36 dependent Spaces on the platform. Developers created 86 finetuned derivatives and 17 adapter models from the base checkpoint.
The broader CodeT5 family includes 761+ models on Hugging Face Hub, based on Hugging Face platform data. That count covers official Salesforce releases, community fine-tunes, and domain-specific adaptations spanning multiple AI research fields.
| Metric | Value |
|---|---|
| Monthly Downloads (CodeT5-base) | 22,172 |
| Community Likes | 132 |
| Dependent Spaces | 36 |
| Finetuned Derivative Models | 86 |
| Adapter Models | 17 |
| Total Models on Hub | 761+ |
Source: Hugging Face platform data, 2026
CodeT5 Model Variants and Parameter Counts
The CodeT5 family spans eight variants. Salesforce released the original three (small, base, large) in 2021, then added five CodeT5+ models in 2023. The newer generation introduced shallow-encoder, deep-decoder architectures for the 2B, 6B, and 16B variants, allowing encoder-only, decoder-only, or full encoder-decoder operation from a single set of weights.
| Model Variant | Parameters | Architecture | Year |
|---|---|---|---|
| CodeT5-small | 60M | Encoder-Decoder | 2021 |
| CodeT5-base | 220M | Encoder-Decoder | 2021 |
| CodeT5-large | 770M | Encoder-Decoder | 2021 |
| CodeT5+ 220M | 220M | Flexible Encoder-Decoder | 2023 |
| CodeT5+ 770M | 770M | Flexible Encoder-Decoder | 2023 |
| CodeT5+ 2B | 2B | Shallow Enc, Deep Dec | 2023 |
| CodeT5+ 6B | 6B | Shallow Enc, Deep Dec | 2023 |
| CodeT5+ 16B | 16B | Shallow Enc, Deep Dec | 2023 |
Source: Salesforce Research, CodeT5+ paper
CodeT5 Model Family: Parameter Scale
CodeT5 Code Generation Accuracy on HumanEval
InstructCodeT5+ 16B hit 35.0% pass@1 on HumanEval zero-shot text-to-code generation, based on the original Salesforce paper. That beat OpenAI’s closed-source code-cushman-001, which scored 33.5%. With CodeT test generation augmentation added, the same model reached 42.9% pass@1 and 67.8% pass@10.
The 770M variant punched well above its weight. At 15.5% pass@1, it performed on par with Incoder 6B (15.2%), GPT-NeoX 20B (15.4%), and PaLM 62B (15.9%) despite having 8 to 80 times fewer parameters. For context, contemporary top models now exceed 90% pass@1 on HumanEval, according to the Papers with Code leaderboard, but these are models with hundreds of billions of parameters or proprietary training data.
| Model | Pass@1 | Pass@10 | Setting |
|---|---|---|---|
| InstructCodeT5+ 16B | 35.0% | 54.5% | Zero-shot |
| InstructCodeT5+ 16B + CodeT | 42.9% | 67.8% | Zero-shot + Test Gen |
| CodeT5+ 770M-py | 15.5% | — | Zero-shot |
| OpenAI code-cushman-001 | 33.5% | — | Zero-shot |
| GPT-NeoX 20B | 15.4% | — | Zero-shot |
| PaLM 62B | 15.9% | — | Zero-shot |
Source: Salesforce Research, CodeT5+ paper; OpenAI Codex paper
HumanEval Pass@1 Comparison (Zero-Shot)
CodeT5 Training Data and Programming Languages
CodeT5+ trained on 51.5 billion tokens from permissively licensed GitHub repositories. That is a 50x jump from the original CodeSearchNet corpus, which contained 8.35 million instances covering six languages. Salesforce restricted the training set to code under MIT, Apache-2.0, BSD-3-Clause, BSD-2-Clause, CC0-1.0, Unlicense, and ISC licenses, based on the model documentation.
CodeT5+ supports nine programming languages. The original release covered eight (Python, Java, JavaScript, Go, Ruby, PHP, C, C#). C++ was added in the CodeT5+ generation. CodeT5-large was pretrained for 150 epochs on the CodeSearchNet corpus.
| Language | CodeT5 (2021) | CodeT5+ (2023) |
|---|---|---|
| Python | Yes | Yes |
| Java | Yes | Yes |
| JavaScript | Yes | Yes |
| Go | Yes | Yes |
| Ruby | Yes | Yes |
| PHP | Yes | Yes |
| C | Yes | Yes |
| C# | Yes | Yes |
| C++ | No | Yes |
Source: Salesforce Research, CodeT5+ documentation
CodeT5 Performance on Math Programming Benchmarks
CodeT5+ 770M hit 87.4% pass@80 on MathQA-Python and 73.8% pass@100 on GSM8K-Python under finetuning, based on the Salesforce paper. Both results outperformed models with up to 137 billion parameters on the same tasks. The sub-billion parameter model’s efficiency on mathematical code generation stands out given its size.
| Benchmark | Model | Result | Setting |
|---|---|---|---|
| MathQA-Python | CodeT5+ 770M | 87.4% pass@80 | Finetuned |
| GSM8K-Python | CodeT5+ 770M | 73.8% pass@100 | Finetuned |
Source: Salesforce Research, CodeT5+ paper
CodeT5 Downstream Task Improvements
CodeT5+ posted gains across 20+ code benchmarks in zero-shot, finetuning, and instruction-tuning settings, according to the Salesforce paper. Text-to-code retrieval improved by an average of +3.2 MRR across eight tasks. Line-level code completion gained +2.1 average exact match across two benchmarks. Retrieval-augmented code generation showed the strongest lift at +5.8 average BLEU-4.
| Task Category | Benchmarks | Improvement |
|---|---|---|
| Text-to-Code Retrieval | 8 tasks | +3.2 avg. MRR |
| Line-Level Code Completion | 2 tasks | +2.1 avg. Exact Match |
| Retrieval-Augmented Code Gen | 2 tasks | +5.8 avg. BLEU-4 |
Source: Salesforce Research, CodeT5+ paper
CodeT5+ Downstream Task Gains
CodeT5 in the AI Code Generation Market
The AI code generation market reached an estimated $8.2 billion in 2025 and is projected to hit $57.6 billion by 2034 at a 24.2% CAGR, according to Dataintelo. GitHub Copilot reported more than 1.8 million paid subscribers in early 2026. Around 92% of developers now use AI tools in some part of their workflow, based on the Stack Overflow 2025 Developer Survey.
CodeT5 sits in the open-source tier of this market. Its Apache 2.0 license makes it free to deploy commercially, unlike proprietary tools from OpenAI or GitHub. The 22,172 monthly downloads are modest against Mistral 7B’s 500,000+ first-month downloads or Whisper’s 4 million monthly pulls, but CodeT5 fills a specific niche in code understanding and generation for teams building on top of smaller, customizable models.
Only 29 to 46% of developers trust AI-generated code outputs, according to the Index.dev developer survey. About 48% of AI-generated code contains potential security flaws, per data from SecondTalent. CodeT5’s encoder-decoder design, which was built for both understanding and generation, gives teams a foundation for tasks like code review automation and defect detection rather than just raw code synthesis.
AI Code Generation Market Size (USD Billions)
CodeT5 Licensing and Repository Status
CodeT5 uses the Apache 2.0 license, allowing unrestricted commercial and research use. Salesforce archived the official GitHub repository in May 2025, based on the repository metadata. Model weights remain fully accessible through Hugging Face, and community forks continue active development.
The open licensing approach separates CodeT5 from closed alternatives like OpenAI’s Codex. Organizations can fine-tune, modify, and deploy CodeT5 models without licensing fees or vendor lock-in. Training produced 49.25 kg of CO2 emissions on Google Cloud Platform, which the provider fully offset through carbon credit programs, according to Salesforce’s published model card.
| Aspect | Detail |
|---|---|
| License | Apache 2.0 |
| GitHub Repo Status | Archived (May 2025) |
| Hugging Face Availability | Active |
| Community Forks | Actively maintained |
| CO2 Emissions (Training) | 49.25 kg (fully offset) |
Source: Salesforce Research, Hugging Face model card
How Does CodeT5 Compare to Other Open-Source Code Models?
CodeT5+ 16B’s 35.0% pass@1 was state-of-the-art for open-source code models at its release. By 2026, newer models have surpassed it. The HumanEval leaderboard’s top score reached 95.1% pass@1 from MiniCPM-SALA, based on the Papers with Code leaderboard. GPT-4o with LDB assistant scored above 90%.
CodeT5’s value in 2026 is less about raw leaderboard position and more about its efficiency and openness. The 770M variant matching models 80 times its size makes it practical for teams with limited compute. The 86 finetuned derivatives on Hugging Face reflect active use for specialized tasks like domain-specific code processing and code summarization, rather than general-purpose chat-style coding assistants.
FAQ
How many monthly downloads does CodeT5 get?
CodeT5-base recorded 22,172 monthly downloads on Hugging Face as of 2026. The total CodeT5 family includes 761+ models on the platform.
What is CodeT5’s accuracy on HumanEval?
InstructCodeT5+ 16B scored 35.0% pass@1 on HumanEval in zero-shot settings. With CodeT augmentation, that rose to 42.9% pass@1.
How many programming languages does CodeT5 support?
CodeT5+ supports nine languages: Python, Java, JavaScript, Go, Ruby, PHP, C, C#, and C++. The original CodeT5 covered eight without C++.
Is CodeT5 free to use commercially?
Yes. CodeT5 uses the Apache 2.0 license, which allows unrestricted commercial deployment, modification, and distribution without fees.
Who developed CodeT5?
Salesforce Research developed CodeT5. The original model released in 2021. CodeT5+ followed in 2023. The GitHub repo was archived in May 2025.
Sources:
https://huggingface.co/Salesforce/codet5-base
https://huggingface.co/Salesforce/codet5p-220m
https://arxiv.org/abs/2305.07922
https://www.index.dev/blog/developer-productivity-statistics-with-ai-tools
