Intel LLM-Scaler-vLLM 1.5 jumps to vLLM 0.26 on Arc GPUs

Intel released version 1.5 of LLM-Scaler-vLLM on October 16, 2026. It is a ready-made Docker image for running large language models on Intel's Arc graphics cards. The update moves its core engine from vLLM 0.14 to vLLM 0.26 and adds newer Qwen and Gemma 4 models. For anyone serving models on Intel GPUs, it closes much of the gap with the code that Nvidia users already run.

vLLM is a popular open-source engine for serving language models. It takes requests, batches them and returns answers quickly. Intel's image bundles a patched version of it with everything else needed to run it on Arc hardware.

What changed in version 1.5

The release moves almost every layer of the stack forward, according to Intel's release notes on GitHub. The image runs on Ubuntu 24.04.4 LTS.

Component	Before	In 1.5
vLLM	0.14.0	0.26.0, with Intel XPU patches
PyTorch	2.10	2.12.0+xpu
Transformers	not listed	5.8.0
Triton XPU	not listed	3.7.1
vllm-xpu-kernels	not listed	0.1.11.2

The jump is large, but it still trails the main project. Upstream vLLM is at 0.31, Phoronix notes. Intel builds the stack under Project Battlematrix, its effort to turn Arc cards into AI workstations. Phoronix describes the release as groundwork for Intel's coming Crescent Island data-center GPU.

The new Docker image is intel/llm-scaler-vllm:1.5. Intel says it is functionally the same as the 0.26.0-b2 beta it published on September 9.

New models and formats

Version 1.5 adds support for five models:

• Qwen3.6-27B
• Qwen3.8-27B
• Qwen3.6-35B-A3B
• gemma-4-26B-A4B-it
• gemma-4-31B-it

It also adds new ways to shrink models. Quantization stores a model's numbers at lower precision so it fits in less memory. The release supports online FP8 and symmetric INT4 quantization, plus checkpoints already saved in block-FP8. It can also keep the KV cache, the memory a model uses to hold the conversation so far, in FP8.

Other additions include LoRA serving and MTP. LoRA adapters are small add-on weights that tune a model for one task. MTP, or multi-token prediction, lets a model propose several tokens at a time. Intel also lists DFlash support, including for Muse-Glimmer-30B.

Intel reports a faster time to first token for Qwen and Gemma 4 models. That is how long a user waits before the answer starts. It also lists faster MTP decoding, LoRA serving and FP8 KV cache.

Bugs fixed and known issues

The release fixes three bugs that produced wrong results:

• Prefix caching did not work correctly with MTP. Prefix caching reuses work when many prompts start the same way.
• Models quantized to symmetric INT4 could return incorrect output.
• Block-FP8 models could corrupt their output when decoding several requests at once.

Intel lists two known issues. "XPU graph support is experimental," the notes say. Multimodal use of Muse-Glimmer also needs Transformers 5.15.0, while the image ships 5.8.0.

What this means for developers

If you serve models on Arc cards with an older LLM-Scaler image, upgrade. The INT4 and block-FP8 fixes alone matter, because both bugs returned wrong answers without crashing.

Check your launch settings against the newer vLLM. Twelve minor versions separate 0.14 and 0.26, and flags or defaults may have changed in between. Test on a staging box before you swap the production image.

If you want Muse-Glimmer with images, wait or build your own image. The bundled Transformers version is too old for it, by Intel's own notes. Keep XPU graph mode off in production while Intel calls it experimental.

The bigger signal is the pace. Intel now ships a vLLM only five versions behind upstream. That makes Arc a more realistic option for local and on-premises model serving.