What are the hardware requirements for Starcoder2 15B?

Starcoder2 15B requires approximately 63.8 GB of VRAM and supports a context window of 16K tokens. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed Starcoder2 15B and how large is it?

Starcoder2 15B is developed by bigcode, a model with 15b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

How does Starcoder2 15B perform on standard benchmarks?

Starcoder2 15B has the following published scores: MMLU 51.35, HumanEval (Python) 44.15. Compare against reference models on this page or on the LLM Explorer leaderboards.

Starcoder2 15B by bigcode — VRAM 63.8GB, 16K context

Name: Starcoder2 15B
Author: bigcode

Starcoder2 15B is an open-source language model by bigcode. Features: 15b LLM, VRAM: 63.8GB, Context: 16K, License: bigcode-openrail-m, LLM Explorer Score: 0.23, Arc: 47.4, HellaSwag: 64.1, MMLU: 51.4, GSM8K: 52.2, HumanEval: 44.2.

Arxiv:2004.05150 Arxiv:2205.14135 Arxiv:2207.14255 Arxiv:2305.13245 Arxiv:2402.19173 Code Dataset:bigcode/the-stack-v2-t... Deploy:azure Endpoints compatible Model-index Region:us Safetensors Sharded Starcoder2 Tensorflow

Model Card on HF 🤗: https://huggingface.co/bigcode/starcoder2-15b

Starcoder2 15B Parameters and Internals

Model Type

text-generation

Use Cases

Areas:

Research, Educational

Applications:

Code generation, Assisting developers

Primary Use Cases:

Text generation based on programming language inputs

Limitations:

May not work with instruction-based commands, Generated code might contain bugs, Not guaranteed to work as intended

Additional Notes

Requires proper attribution when utilizing generated code from the model due to licenses.

Supported Languages

600+ programming languages ()

Training Details

Data Sources:

The Stack v2, Arxiv, Wikipedia

Data Volume:

4+ trillion tokens

Methodology:

Grouped Query Attention, Sliding Window Attention, Fill-in-the-Middle objective

Context Length:

16384

Hardware Used:

NVIDIA DGX H100, 1024 x H100 GPUs

Model Architecture:

Transformer decoder

Input Output

Input Format:

Text prompt

Output Format:

Generated text/code

LLM Name	Starcoder2 15B
Repository 🤗	https://huggingface.co/bigcode/starcoder2-15b
Model Size	15b
Required VRAM	63.8 GB
Updated	2026-07-25
Maintainer	bigcode
Model Type	starcoder2
Model Files	4.6 GB: 1-of-14 4.6 GB: 2-of-14 4.6 GB: 3-of-14 4.6 GB: 4-of-14 4.6 GB: 5-of-14 4.6 GB: 6-of-14 4.6 GB: 7-of-14 4.6 GB: 8-of-14 4.6 GB: 9-of-14 4.6 GB: 10-of-14 4.6 GB: 11-of-14 4.6 GB: 12-of-14 4.6 GB: 13-of-14 4.0 GB: 14-of-14
Model Architecture	Starcoder2ForCausalLM
License	bigcode-openrail-m
Context Length	16384
Model Max Length	16384
Transformers Version	4.37.0.dev0
Tokenizer Class	GPT2Tokenizer
Vocabulary Size	49152
Torch Data Type	float32

Rank the Starcoder2 15B capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Quantized Models of the Starcoder2 15B

Model	Likes	Downloads	VRAM
Starcoder2 15B AWQ	1	8	9 GB
StarCoder2 15B GGUF	31	1731	6 GB

Best Alternatives to Starcoder2 15B

Best Alternatives	Context / RAM	Downloads	Likes
Starcoder2 15B Instruct V0.1	16K / 31.9 GB	597	105
Starcoder2 15B Finetuned Drake	16K / 63.8 GB	7	0
CodeFuse StarCoder2 15B	16K / 31.9 GB	18	1
Starchat2 15B V0.1	16K / 31.9 GB	352	112
Starcoder2 15B Instruct V0.1	16K / 53.4 GB	25	0
...aceH4 Starchat2 15B V0.1 4bits	16K / 9.9 GB	7	0
Dolphincoder Starcoder2 15B	16K / 31.9 GB	84	69
Dolphincoder Starcoder2 15B	16K / 31.9 GB	17	68
Starchat2 15B Sft V0.1	16K / 31.9 GB	12	5
Starcoder2 15B Instruct	16K / 31.9 GB	6	7