StarCoder2-15B, the largest open-access code generation model from the BigCode project, was trained on over 4 trillion tokens across 619 programming languages. It matches or outperforms models more than twice its size on several benchmarks, according to the project’s February 2024 paper. This article covers StarCoder statistics for 2026, including model specs, benchmark performance, training data scope, developer demographics in AI-assisted coding, and how StarCoder compares to other open-source code LLMs.
StarCoder Statistics 2026 — TL;DR
The BigCode collaboration has grown to over 1,200 members from 62 countries, based on a ServiceNow governance report.
StarCoder2-15B scored 46.3% pass@1 on HumanEval+ with greedy decoding, outperforming CodeLlama-13B by 17 percentage points.
The Stack v2, used to train StarCoder2, spans 619 programming languages and is 4x the size of its predecessor.
StarCoder2-3B, the smallest variant at 3 billion parameters, matches StarCoder1-15B in coding benchmarks despite being 5x smaller.
84% of developers now use or plan to use AI coding tools in their workflow, according to the Stack Overflow 2025 Developer Survey of 49,000+ respondents. StarCoder remains one of the few fully transparent alternatives: open weights, open training data, and a published opt-out process.
How Many Parameters Does StarCoder Have?
The StarCoder family includes two generations. StarCoder1, released in May 2023, is a single 15.5 billion parameter model. StarCoder2, released in February 2024, comes in three sizes: 3B, 7B, and 15B.
All StarCoder2 models use Grouped Query Attention and a 16,384-token context window with sliding window attention of 4,096 tokens. StarCoder1 used Multi Query Attention and an 8,192-token context window.
| Model | Parameters | Context Window | Training Tokens | Languages |
|---|---|---|---|---|
| StarCoder1 | 15.5B | 8,192 | 1T | 80+ |
| StarCoder2-3B | 3B | 16,384 | 3.3T | 17 |
| StarCoder2-7B | 7B | 16,384 | 3.5T | 17 |
| StarCoder2-15B | 15B | 16,384 | 4.3T | 619 |
Source: Hugging Face / BigCode Project
StarCoder Coding Performance on HumanEval
HumanEval is the most widely cited benchmark for code generation. It consists of 164 hand-written Python programming tasks. Pass@1 measures whether the model’s first attempt passes all unit tests.
StarCoder1 scored 33.6% pass@1 on HumanEval in its base configuration, rising to about 40% with optimized prompting. StarCoder2-15B improved on that baseline and outperformed CodeLlama-34B on both MBPP and MBPP+, despite having less than half the parameters.
| Model | HumanEval (pass@1) | Parameters |
|---|---|---|
| StarCoder1 (base) | 33.6% | 15.5B |
| StarCoder1 (optimized) | ~40% | 15.5B |
| StarCoder2-15B | 46.3% | 15B |
| CodeLlama-34B | 48.8% | 34B |
| DeepSeek-Coder-33B | 47.6% | 33B |
| Qwen2.5-Coder-32B | 92.7% | 32B |
Source: BigCode paper (OpenReview), Qwen team, DeepSeek paper
Qwen2.5-Coder-32B leads the open-source field at 92.7%, but it was released over a year after StarCoder2. Within the early-2024 cohort, StarCoder2-15B held its own against models with double the parameters.
StarCoder Training Data — The Stack v1 and v2
StarCoder1 was trained on The Stack v1.2, a dataset of 783GB of permissively licensed code from GitHub. It covered 86 programming languages and included 54GB of GitHub issues plus 13GB of Jupyter notebooks. The total was around 250 billion tokens.
StarCoder2 used The Stack v2, built in partnership with Software Heritage. The Stack v2 is 4x larger than v1, spanning 619 programming languages and pulling from GitHub pull requests, Kaggle notebooks, and code documentation alongside the Software Heritage archive.
| Dataset | Size | Languages | Sources |
|---|---|---|---|
| The Stack v1.2 | 783 GB | 86 | GitHub repos (permissive licenses) |
| The Stack v2 | ~3.1 TB | 619 | Software Heritage, GitHub PRs, Kaggle, docs |
Source: Hugging Face Datasets / BigCode
Who Built StarCoder? BigCode Developer Demographics
BigCode is an open scientific collaboration co-led by Hugging Face and ServiceNow. As of mid-2024, the project had over 1,200 members from institutions and companies across 62 countries, based on a ServiceNow case study. The original StarCoder paper listed more than 600 contributing researchers.
StarCoder2-3B was trained by ServiceNow. StarCoder2-7B was trained by Hugging Face. StarCoder2-15B was trained by NVIDIA using NeMo on the Eos Supercomputer with 1,024 H100 GPUs. The three organizations split responsibility for different model sizes.
On GitHub, the bigcode-project/starcoder repository has 7,500 stars and 530+ forks. The StarCoder2 code generation models are distributed through Hugging Face Hub, where the starcoder2-3b variant alone has over 202,000 downloads.
StarCoder2 Hugging Face Downloads
| Model Variant | Downloads (as of early 2026) | Hugging Face Likes |
|---|---|---|
| StarCoder2-3B | ~202,000 | 209 |
| StarCoder2-7B | ~20,100 | 198 |
| StarCoder2-15B | ~4,660 | 640 |
Source: Hugging Face Hub (bigcode collection page)
The 3B model accounts for the bulk of downloads. Smaller models are easier to run on consumer-grade hardware, and StarCoder2-3B already matches StarCoder1-15B on most benchmarks.
StarCoder vs Other Open-Source Code LLMs
The open-source code LLM space has moved quickly since StarCoder2’s release. Qwen2.5-Coder-32B, released in late 2024, hit 92.7% on HumanEval. DeepSeek-Coder-V2, a 236B mixture-of-experts model under an MIT license, is another strong competitor. Code Llama from Meta offers sizes from 7B to 70B with solid Python performance.
StarCoder2 still stands out in two areas. First, it covers 619 programming languages, far more than most competitors. DeepSeek Coder and Qwen Coder focus on fewer high-resource languages. Second, BigCode publishes every preprocessing script, opt-out policy, and training configuration. For teams that need full audit trails, that transparency matters.
| Model | Parameters | Languages | License | Training Data Open? |
|---|---|---|---|---|
| StarCoder2-15B | 15B | 619 | OpenRAIL-M | Yes |
| Code Llama 34B | 34B | ~15 | Custom (Llama) | No |
| DeepSeek-Coder-33B | 33B | ~90 | Custom | No |
| Qwen2.5-Coder-32B | 32B | ~90 | Apache 2.0 | No |
Source: Respective model documentation on Hugging Face
A March 2026 review from CheckThat.ai noted that StarCoder’s 33.6% HumanEval base score places it behind commercial models, but its permissive licensing and training data transparency fill a gap that most closed-source tools cannot.
How Many Developers Use AI Coding Tools in 2026?
According to the JetBrains 2025 Developer Ecosystem Survey of 24,534 developers, 85% regularly use AI tools for coding. The Stack Overflow 2025 Developer Survey (49,000+ respondents) put the figure at 84% using or planning to use AI tools, up from 76% in 2024.
51% of professional developers now use AI tools every single day, per Stack Overflow. A McKinsey study from February 2026, which surveyed 4,500 developers across 150 enterprises, found that AI tools cut time on routine coding by an average of 46%.
GitHub Copilot reached 20 million cumulative users by mid-2025 and had 4.7 million paid subscribers by January 2026. Open-source alternatives like StarCoder, Continue, and Tabby offer a zero-cost path for developers who need data privacy or want to avoid vendor lock-in.
Trust in AI coding output is falling. Only 29% of developers trust AI tool output, down from 40% in 2024, according to Stack Overflow. That gap between high adoption and low trust defines the current state of the market.
StarCoder Programming Language Coverage
StarCoder1 was trained on 86 programming languages. StarCoder2-15B expanded that to 619. The smaller 3B and 7B variants of StarCoder2 cover 17 languages.
The 17 languages covered by the smaller models include Python, JavaScript, TypeScript, Java, C, C++, C#, Go, Rust, Ruby, PHP, Swift, Kotlin, Scala, Lua, R, and Shell. StarCoder2-15B adds hundreds of lower-resource languages, domain-specific languages, and configuration formats from the Software Heritage archive.
On the MultiPL-E benchmark, which tests code completion across 18 languages, StarCoder2-15B matched or exceeded CodeLlama-34B on 10 of those 18. It outperformed DeepSeek-Coder-33B specifically on lower-resource languages like D, Julia, Lua, and Perl.
StarCoder Licensing and Commercial Use
All StarCoder models are released under the BigCode OpenRAIL-M v1 license. This allows commercial use, modification, and redistribution, provided users follow certain responsible AI conditions. The license is more permissive than Meta’s Llama license, which requires branding obligations, and more restrictive than Apache 2.0 or MIT.
The training data is fully disclosed. Every source code file used in training can be traced through Software Heritage Persistent Identifiers (SWHIDs). Developers can also opt out of having their code included in future training runs, a feature few other code LLMs offer.
StarCoder Infrastructure and Compute Costs
StarCoder2-15B was trained on 1,024 NVIDIA H100 GPUs using the Eos Supercomputer and NeMo framework. The estimated total carbon emissions for training came to about 16,107 kg CO2eq, based on a carbon efficiency of 0.2925 kgCO2eq/kWh at Hugging Face’s compute infrastructure.
Running StarCoder2-15B in FP16 requires roughly 30GB of GPU memory. In 8-bit quantization, it fits under 20GB. The 3B variant uses about 4.2GB in 4-bit mode, making it runnable on consumer GPUs with 6GB VRAM.
| Model | FP16 Memory | 8-bit Memory | 4-bit Memory |
|---|---|---|---|
| StarCoder2-3B | ~6 GB | ~4.5 GB | ~4.2 GB |
| StarCoder2-7B | ~14 GB | ~8.5 GB | ~4.2 GB |
| StarCoder2-15B | ~30 GB | ~18 GB | ~10 GB |
Source: Hugging Face model cards / BigCode GitHub
FAQ
What is StarCoder used for?
StarCoder is an open-source code generation model for tasks like code completion, code editing via fill-in-the-middle, and acting as a technical assistant when prompted correctly. It supports 619 programming languages in its largest variant.
Is StarCoder free to use commercially?
Yes. StarCoder is released under the BigCode OpenRAIL-M v1 license, which permits commercial use, modification, and redistribution with responsible AI conditions attached.
How does StarCoder2 compare to GitHub Copilot?
GitHub Copilot uses proprietary models from OpenAI. StarCoder2-15B scores lower on HumanEval but offers full transparency of training data, open weights, and no subscription cost for self-hosted deployments.
What hardware do I need to run StarCoder2?
StarCoder2-3B runs on a GPU with 6GB VRAM in FP16. The 15B model needs about 30GB in FP16, or under 20GB in 8-bit quantization. Consumer GPUs like the RTX 4090 can handle the 15B variant.
Who develops StarCoder?
StarCoder is built by the BigCode project, an open collaboration led by Hugging Face and ServiceNow with over 1,200 members from 62 countries. NVIDIA trained the 15B variant of StarCoder2.
Sources:
https://huggingface.co/blog/starcoder2
https://arxiv.org/abs/2402.19173
https://www.servicenow.com/blogs/2024/bigcode-open-innovation-case-study
