Qwen3.5 27B GGUF picks hinge on eval rigor

Hot take: there is no universal “best GGUF” here yet; the winner depends on whether you optimize for raw coding accuracy, instruction reliability, or throughput at your exact context length.

• –Thread consensus is still anecdotal, but Unsloth UD quants keep coming up because they publish quantization methodology and updated calibration notes.
• –Distilled/finetuned packs (for example Claude-distilled variants) can outperform base quants on some coding prompts, but they should be compared as a different model recipe, not just “better quantization.”
• –A fair comparison should lock prompt set, seeds, context window, backend (llama.cpp/Ollama/LM Studio), KV cache precision, and then track pass@1 plus compile/test success, not only tokens/sec.
• –KLD/perplexity are useful screening signals, but practical coding quality often diverges, so include real repo tasks (bug fix, refactor, multi-file edit) in your eval harness.
• –For a 20–24GB target, Q4_K_M vs Q4_K_XL vs Q5_K_M trade-offs are usually the key decision point: Q5 tends to improve consistency, Q4 tends to improve speed and fit.