Repos made with Qwen3.8-Flash-Next

Device-side LRU expert cache + kernel-count patches: Qwen3.8-Flash-Next decode on 2x AMD R9700 (gfx1201), ROCm 10, tcclaviger vLLM fork

Serving Qwen3.8-Flash-Next (Qwen4 preview) on 2x RTX PRO 6000 Blackwell SM120 with vLLM: FP8 & NVFP4, MTP spec decoding, PLE offload, 262k context. Recipe, pitfalls, benchmarks.

Measured llama.cpp tuning for Qwen3.8-Flash-Next (UD-IQ4_XS) on a 16GB consumer GPU — speculative-decoding variant sweep, expert-residency slope, context scaling, and negative results

Qwen3.8-Flash-Next W4A16 on 2x CMP 170HX with vLLM: ~188 tok/s, 1M KV tokens, FP8 PLE on a 91 GB host

vLLM patches + root-cause forensics: Qwen3.8-Flash-Next with PP3, MTP, and prefix caching (fixes the vllm#54173-family state corruption)

Quantization benchmark: Qwen3.8-Flash-Next GGUF quants vs code generation quality

Qwen3.8-27B + Flash-Next benchmarks on dual RTX 3090: Chain-of-Draft prompt takes GSM8K from 23% to 77%

Qwen3.8-Flash-Next (176B MoE) auf drei gebrauchten RTX 3090: Messwerte, Harness und Ergebnisse

Low-memory Apple Silicon runtime for local LLM inference, focused on Qwen models and Flash-Next on 16 GB Macs.