Qwen3.8-27B

Proceed with caution

Why

Official benchmarks (SWE-bench Pro 61.7 vs Qwen3.7-Plus's 57.6; OSWorld-Verified 84.3 vs 73.3) are strong and the ~24GB consumer-VRAM local-deployment claim is independently corroborated by community GGUF quantization data, but the model is only ~3 days old and early-adopter threads already report excessive-reasoning stalls (49+ min), local-inference crashes, and quality regressions vs. the prior generation — real open questions before broad reliance.

Next

Pilot on non-critical, sandboxed coding-assistant workloads using the Q4_K_M GGUF quant (~17.8GB, fits in 24GB VRAM) or the official FP8 build; hold off on production-critical agentic use until reasoning-loop stalls and llama.cpp crash reports are resolved in a patch release; re-check in 2-4 weeks once community stability reports and independent (non-vendor) benchmarks mature.