Use cases
coding
Summary
User compares Strata running Qwen3.8 Flash-Next at 3-bit against llama.cpp running Qwen3.8 27B Q8 with MTP on two R9700 cards (64 GB total).
Strata was stable only with all experts in VRAM, which on 64 GB meant a 3-bit quant; offloading experts to RAM hung on long tasks. The user reports Strata was faster on paper, around 60 t/s versus llama.cpp.
On HumanEval+ the 3-bit Flash-Next scored 93.9% against 95.1% for Qwen3.8-27B Q8 on llama.cpp, and needed nearly twice the tokens. The user sees no real-world gain over a mature stack and asks whether model-specific engines are overhyped.