Baekpica/Qwen3.8-Flash-Next-Mixed-Quant-SSD-PLE-GGUF
Note rejected QSA PV follow-up; no new default path
Pin Qwen performance report to merged runtime
Document measured MTP runtime optimizations
Document FP8 PLE runtime support and paired 64K sweep
Add official FP8 PLE sidecars and download examples
Update README.md
Document measured image latency and agent checks
Model card: 2026-09-07 prefill rounds on ds4-dfm-rs (6e036c4), refreshed 2K-64K graph
Update README.md
Refresh 2K-64K sweep after D2R r4
Record the 2026-09-06 ds4-dfm-rs prefill rounds
Refresh long-context MTP throughput
docs: add long-context throughput chart
Record the 2026-09-02 ds4-dfm-rs prefill campaign (908.1 tok/s at 8,259 tokens)
Document verified 512K YaRN boundary
docs: publish Rust-host DGX Spark results
Add GitHub Sponsors button
Update README.md
Add support link
Restore Qwen model card and latest validation
Document full-context SSD-PLE latency profile
Document guarded full-window Qwen validation
Document two-bank router projection
Document two-bank GDN projections
Document two-bank router top-k
Document two-bank Q5_K routed main
Document two-bank QSA projection
Document paired two-bank Q5 tail
Document two-bank SSD-PLE gather
docs: correct Q8 output-head parity width
docs: record two-bank Q8 output batching
Document paired two-bank GDN recurrent update
Document continuous bank reasoning continuation
Document state-safe Qwen row batching
Document tiled large-image vision attention
Document bounded PLE cache and worker tuning
Document decoded-pixel image cache reuse
Refresh model card performance highlights
Document verified still-image input
Document guarded Qwen still-image serving
docs: record guarded Q8 optimization rounds
docs: record two-bank embedded MTP verification
Document 262K API and serial fallback checks
docs: record Qwen reuse MTP and disk KV verification
docs: record latest Qwen prefill optimizations
docs: record Q8 PLE split performance
Document latest QSA prefill optimization
Document optimized Q5 serving performance
docs: record Q5 DGX Spark serving profile
docs: record 262K demand-mapped serving smoke