Intel software engineers today released the latest version of LLM-Scaler-vLLM PV 1.5, their containerized Docker stack providing a convenient AI/LLM experience atop Intel Arc Graphics hardware and born out of their Project Battlematrix initiative.

This is a significant update to their packaged vLLM software stack for making use of Arc Graphics and ahead of the Crescent Island Intel GPU D Series. With this v1.5 release, they jumped all the way from vLLM 0.14 to vLLM 0.26. Quite a version leap and putting them closer to the upstream state. Albeit still behind with vLLM currently at version 0.31.

In addition to the vLLM 0.14 to 0.26 leap, PyTorch jumped from v2.10 to v2.12 with this Docker setup. The LLM coverage has also improved with more Qwen models now working (Qwen3.6-27B, Qwen3.8-27B, and Qwen3.6-35B-A3B), support for more Gemma models (gemma-4-26B-A4B-it and gemma-4-31B-it), and also now supporting online FP8 and symmetric INT4 quantization. FP8 KV cache is also now supported.

The Intel software stack has also seen performance improvements around faster time-to-first-token for Qwen and Gemma 4 models, improved MTP decoding performance, and improved LoRA serving and FP8 KV cache performance.

Bug fixes round out this new Intel software release that can be found via GitHub.