sergiiob/intel-arc-pro-b70-inference-cookbook released version 1.11.0 on November 5, 2026. It is a repository collecting LLM inference recipes, engine patches, and a benchmark harness for Intel Arc Pro B60/B70 (Xe2, Battlemage), and what changed in this release is twofold: attaching a lossy notation to the SC5 entries in the quantization catalog, and adding an explanation of the parity column to the README.
The gap this repository fills is covering, with recipes and patches, the parts that cannot be reproduced with published settings alone. The README cites, on a single-stream basis for MoE 35B, 160 t/s decode and 7.5K t/s prefill, but these are figures the repository itself discloses, not independent test results. The catalog records not only what each recipe runs and how fast, but also how far its output differs from the reference. The parity column is where that comparison lives, and SC5 is a notation referring to 5.00 bits per weight (bpw). 1.11.0 explained the meaning of this column in the README and left the lossy notation on the SC5 row 241/329, while the record of 1.10.0, released on the same day, lists the KL value of Qwen3.8-27B SC5 as 0.418x and 0.161x of ordinary 4.00bpw.
That said, 1.11.0 itself contains no newly measured performance figures. It is an automated release commit in which only the version in the manifest changed from 1.10.0 to 1.11.0 (release PR #70), and the actual content is the catalog and README modifications that came in through issue #69. Because the commit description says neither on what basis 241/329 was counted nor what the parity column compares against what, it is difficult to gauge the magnitude of the loss from the notation alone. The fact that so many entries carry the lossy notation merely creates a reason to check the parity values of those rows first if you plan to run a 35B- or 27B-class model on Arc Pro B70; it does not indicate any performance improvement. It is better read as a cleanup of notation that narrows the choices, rather than a notable feature addition.