Skip to content

2026.11.17.01

Choose a tag to compare

@rizagp rizagp released this 17 Nov 02:24

Windows: unpack inferzero-2026.11.17.01-e9ff57f-win64.zip anywhere, then run setup.cmd (or zero-launcher.cmd once a model is there).

Linux: inferzero-2026.11.17.01-e9ff57f-linux-x64.tar.gz (Ubuntu 24.04 and newer): unpack it, then ./zero-launcher.sh --check and ./zero-launcher.sh. It needs Intel's GPU runtime and nothing else. The steps: docs/install.md.

New since 2026.11.16.16:

  • Panther Lake (Arc B390) and Lunar Lake: about twice the prompt speed on dense Qwen models. These iGPUs now take the discrete cards' prompt path (H8 on the wide-tile XMX GEMM, and flash attention). A B390's Qwen3.8 27B 4K prompt went from ~260 to ~550 tok/s (issue #1); accuracy checked (KL 0.0018). ARC_H8=0 / ARC_ATTN_FX=0 turn them off.
  • Flash Next's MTP head in Q8_0 loads (mtp-Qwen3.8-Flash-Next-Q8_0.gguf): it stopped with "unexpected expert tensor blk.48..." (issue #31).
  • serverz --mtp-model FILE, as in runz: an MTP head or drafter in a file of its own. After a switch to another model the head is skipped with a note.
  • Two test switches for discrete cards' MoE prompts, off by default until a card measures them; both give the same results as without:
    • ARC_MOE_RG_XMX=1: the expert GEMMs on XMX straight from the expert rows, without the per-chunk decode pass (the 35B and other MoE models; Flash Next's down projection).
    • ARC_MOE_GEMM_PAD=1: Flash Next's expert gate / up through decode + the XMX GEMM instead of the per-tile kernel.
  • Smaller packages: the developer notes and the test PDFs are no longer included (they stay on GitHub).