Intel LLM-Scaler vLLM 1.5: Updated inference engine, broader model and quantization support

1. The underlying engine and image

Released on October 16, 2026, this version upgrades the underlying engine from vLLM 0.14.0 to vLLM 0.26.0. The configuration includes patches for Intel GPUs.

The distributed image is intel/llm-scaler-vllm:1.5. The release notes describe it as functionally equivalent to intel/llm-scaler-vllm:0.26.0-b2. There is therefore no basis for treating the tags as different feature generations.

2. Supported models and quantization paths

The release notes list Qwen3.6-27B and Qwen3.8-27B as supported models. Both also support files with weights stored at lower precision in blocks.

Qwen3.6-35B-A3B is included in the same support scope. All three models can use pre-quantized FP8 weights.

For the Gemma family, gemma-4-26B-A4B-it and gemma-4-31B-it are listed. The reference documentation also specifies runtime quantization support for both models.

Online quantization reduces numerical precision at runtime instead of using separately processed weight files. This release supports FP8, an 8-bit floating-point format, and INT4, a symmetric 4-bit integer format. For pre-quantized models, the instructions say to pass the model path without specifying quantization options again.

3. Response generation and customized model serving

MTP is a speculative decoding method that predicts multiple tokens, the pieces that make up sentences, in advance. The Qwen family uses the model's own prediction capability.

DFlash also provides speculative decoding by proposing and verifying multiple tokens in advance. The release notes list the three Qwen models above and Muse-Glimmer-30B as supported. The reference documentation advises keeping other server and measurement settings identical when comparing runs with and without it.

LoRA serves fine-tuned variants attached to a single base model. Multiple adapters containing trained variants can be served together and loaded or unloaded at runtime. The release notes specify support for the five Qwen and Gemma models listed above.

FP8 KV cache stores the keys and values of intermediate computations reused in response generation at lower precision. It is configured independently of weight quantization, so both can be used together. However, runtime cache scale calculation is unsupported for some model and execution environment combinations.

Benefits include running without pre-quantized files and serving multiple variants from one base model. However, Gemma multi-token prediction and DFlash require separate auxiliary model files suited to the target.

Qwen 3.8, listed as supported in the release notes, is absent from the reference documentation's validation lists for multi-token prediction, adapters, and auxiliary-model acceleration. Support declarations should be distinguished from specific execution guidance.

4. Improvement claims and verification scope

Intel says it has improved time to first token for Qwen and Gemma models. Multi-token prediction, adapter serving, and low-precision cache performance are also listed as improvements. However, no measurement conditions or figures showing the gains were provided for these items.

Fixes include an error when combining multi-token prediction with prefix caching. Prefix caching reuses computations for shared input prefixes. Incorrect output from symmetric 4-bit integer quantized models was also fixed. Another fix addresses corrupted output during concurrent decoding with block-quantized models. This list should not be interpreted as an accuracy guarantee for every combination.

5. Requirements to check before installation

In its image guidance for Arc B60, the reference documentation recommends downloading an exact version rather than an unpinned latest tag. Below is the original download example. Replace the placeholder with the release version to use.

# Replace <VERSION> with the latest beta release version from the link above
docker pull intel/llm-scaler-vllm:<VERSION>

For production, the documentation advises building a custom image using the published Dockerfile as a reference.

The server will not start if the maximum context length configured at startup exceeds available cache capacity. Check the startup logs and reduce the length or adjust memory usage settings.

Muse-Glimmer inference that also processes images and video requires Transformers 5.15.0, but the image contains Transformers 5.8.0.

This version broadens local inference options, but adoption decisions should rest on model-specific execution requirements and testing under identical conditions rather than feature lists.