Revision History

Verification fixes: Splash memory plan computed at startup; citations; Inco DFlash claim attributed

Verification fix: z-lab DFlash 2 repo predates the incoai one, so it is not a mirror

Add Apple silicon section: Inco AI Splash engine with DFlash 2 draft for Qwen3.8-27B (Inco M5 Pro measurements, 144 tok/s M5 Max provenance); link DFlash and Inco AI

Added NVIDIA Jetson subsection: day-zero Jetson support, Jetson AI Lab platform list, the Inferact NVFP4 and Inco AI DFlash2 community checkpoints with publishers labeled, NVIDIA's Sep 4, 2026 speedups and SpeedBench ranges with the module caveat, the vLLM recipe, and the attributed chart placement; added Flash-Next and Max-0902 timeline entries and a dated Artificial Analysis Index v4.2 note

Added bounded local deployment measurements and a two-instance demonstration with methodology caveats.

Updated Qwen3.8 with exact Dynamic V3 GGUF metadata, corrected token-agreement claims, dated memory guidance, and evaluation limits.

Updated the Qwen3.8 article with a dated independent benchmark snapshot, current low-bit memory guidance, and clearer limits on single-GPU performance claims.

Expanded the Qwen3.8 article with verified Qwen3.8-27B release chronology, chat-template reasoning semantics, framework requirements, and qualified hardware-specific runtime measurements.

Update Qwen3.8 with the verified August open-weight releases, separating the 2.4T-A95B and 27B checkpoints and documenting their architecture, modalities, context, licenses, artifacts, and qualified benchmark and hardware details.

Qwen3.8-Max released August 3, 2026: linked new article, updated preview framing, noted weights still pending as of August 5

No summary provided