StarCoder 2 15B from BigCode is a permissively-licensed coding model with particularly broad language support – over 600 programming languages in its training data. On our dedicated GPU hosting it fits a 16-24 GB card with practical throughput.
Contents
VRAM
| Precision | Weights | Fits On |
|---|---|---|
| FP16 | ~30 GB | 32 GB+ card |
| FP8 | ~15 GB | 16 GB card tight, 24 GB comfortable |
| AWQ INT4 | ~9 GB | 12 GB+ card |
Deployment
python -m vllm.entrypoints.openai.api_server \
--model bigcode/starcoder2-15b-instruct-v0.1 \
--quantization awq \
--max-model-len 16384 \
--gpu-memory-utilization 0.92 \
--enable-prefix-cachingStarCoder supports fill-in-middle via special tokens. See the model card for the exact FIM format.
Licence
StarCoder 2 ships under the BigCode OpenRAIL-M licence – more permissive than Meta’s Llama licence and allows commercial use with ethics restrictions. For teams needing a clean commercial licence story on dedicated hosting, StarCoder 2 is often preferable to Qwen Coder or Codestral on licence grounds alone.
Versus Alternatives
| Model | Quality | Licence |
|---|---|---|
| StarCoder 2 15B | Good | OpenRAIL-M (permissive) |
| Codestral 22B | Better | Mistral Non-Production (restrictive) |
| Qwen Coder 32B | Best | Qwen licence |
Permissively-Licensed Code AI
StarCoder 2 on UK dedicated hosting – clean licence story for commercial deployments.
Browse GPU ServersSee Qwen Coder 32B and Codestral 22B.
