Three small-to-mid code models — one firmly current, one holding niche value, one largely obsolete for production. Qwen 2.5 Coder 7B Instruct was released by Alibaba in November 2024 and is the clear benchmark leader of this trio. At 7 billion parameters it competes directly with DeepSeek Coder 6.7B on HumanEval while offering a 131K context window that is unusually large for its size class. That context depth matters for IDE autocomplete sessions where you want to pass entire file trees as context without chunking. Hosted pricing is typically below $0.20 per million tokens, making high-frequency completions economically viable. The Qwen license permits commercial use. StarCoder2 15B Instruct, released September 2024 by the BigCode collaboration between HuggingFace and ServiceNow, runs at 15 billion parameters with a 16K context window. It trails Qwen 2.5 Coder on raw HumanEval, so raw benchmark performance is not its pitch. What distinguishes it is training-data provenance: every training sample comes from The Stack v2, a dataset restricted to permissively licensed open-source code. Teams in regulated industries or those with IP audit requirements often prefer StarCoder2 specifically because the training-data chain is verifiable. The BigCode OpenRAIL-M license is commercially usable with narrow application restrictions. Stable Code Instruct 3B was released by Stability AI in January 2024 and covered single-file fill-in-middle tasks at 3B parameters with a 16K context window. By mid-2026 it has been largely displaced: Qwen 2.5 Coder 7B and DeepSeek Coder deliver materially better HumanEval scores at comparable hosted cost. Stability AI's non-commercial community license also adds friction — commercial use requires a Stability membership — making it harder to justify against freely commercial alternatives. Pick Qwen 2.5 Coder 7B for production autocomplete. Pick StarCoder2 15B when training-data provenance is a compliance requirement. Skip Stable Code 3B for production work unless you are specifically targeting constrained on-device inference with legacy hardware limits.