Qwen2.5-Coder sparks Swift coding debate

Qwen2.5-Coder looks like the right first local model to try, but Swift is one of those languages where generic code bragging rights do not settle the argument. The more honest read is that it is a strong contender, not a proven king.

• –Qwen says Qwen2.5-Coder is trained on 5.5T code tokens, supports 92 programming languages including Swift, and ships in 0.5B to 32B sizes, so the family is genuinely built for code: https://qwen2.org/qwen2-5-coder/ and https://qwenlm.github.io/blog/qwen2.5-coder-family/
• –SwiftEval is a Swift-specific benchmark with 28 hand-crafted problems, and its paper puts Qwen2.5-Coder Instruct 32B at 79.1, which is strong but not a runaway win: https://swifteval.macpaw.com/ and https://research.macpaw.com/publication-assets/swift-eval/SwiftEval__Industry_Experience_in_Developing_a_Language_Specific_Benchmark_for_LLM_generated_Code_Evaluation__ICSE_25_FORGE_Data_and_Benchmarking.pdf
• –The smaller Qwen2.5-Coder variants trail much more on SwiftEval, so local-only setups need to choose size carefully instead of assuming the family is uniformly strong.
• –The Claude/Codex plus local-model split is a sensible cost-saving workflow, because the local model can chew through boilerplate, scaffolding, and repetitive edits while the frontier model audits correctness.
• –For Swift, language-specific benchmarks matter more than Python-centric leaderboards, because translated tasks can hide Swift-specific failure modes.