LLM Scaler: Intel's Arc Pro Docker Line Has Missed Exactly One Month in Fifteen

By Indie Kings | October 19, 2026

Updated October 19, 2026: Intel maintains LLM Scaler, an Apache-2.0 serving stack that makes vLLM, SGLang, ComfyUI and Xinference run on Arc Pro B60 and B70 GPUs, and the repository was last pushed October 17, 2026. In the most recent week it shipped explicit cache-consistent GDN, Q8 and fused MoE APIs for SGLang (October 17), an Omni 0.2.0-b2 version bump (October 16), and packaged LMCache KVWeave cache offload for the vLLM image (October 14), all per the repository commit history.

Image credit: intel/GitHub.

What LLM Scaler actually is

Per the official repository README, LLM Scaler is a GenAI solution for text generation, image generation and video generation running on Intel Arc Pro B60 and B70 GPUs. It leverages standard frameworks, named by Intel as vLLM, ComfyUI, SGLang Diffusion and Xinference, and Intel says it tunes those stacks for the best performance of state-of-the-art GenAI models on those two cards.

The unit of delivery is a Docker image, not a driver and not a fork of the frameworks. Every release listed on the repository's Releases page is a Docker tag. The current vLLM image is tagged intel/llm-scaler-vllm:0.26.0-b2, and the Releases page separately lists nine tags across the Omni image line, from 0.1.0-b1 in September 2025 through 0.2.0-b1 in August 2026.

The repository itself is an Intel GitHub org project written mostly in C++, per the repository metadata, created June 19, 2025 and pushed October 17, 2026. At an October 18, 2026 fetch it showed 568 stars, 86 forks and 180 open issues, under the Apache-2.0 license.

What moved this week

Three changes landed between October 14 and October 17, 2026, each verified directly in the commit history.

The most recent commit, dated October 17, is titled "sglang: add explicit cache-consistent GDN, Q8 and fused MoE APIs (#763)". Per its commit body, it makes GDN page-state rounding an explicit backward-compatible API and adds opt-in batch-invariant Q8 GEMV and fused MoE decode companions. In plain terms, MoE means mixture-of-experts decoding, GEMV means matrix-vector multiply, and Q8 is 8-bit quantization, so the commit gives SGLang users explicit control over cache-consistent checkpoints and quantized decode paths. The commit is cryptographically signed and GitHub reports the signature valid.

Two commits dated October 16 handle the Omni image: "Omni: bump 0.2.0-b2 version" and "build(omni): mark 0.2.0-b2 release". The version bump is verified, but the README and Releases.md had not documented what 0.2.0-b2 contains beyond the bump at fetch time, so any claim about its feature list would be unverified and this article makes none.

The third change, dated October 14, is "vllm: package LMCache KVWeave cache offload (#707)". LMCache offloads key-value cache state out of GPU memory, and the vLLM section of the README already points readers to a Qwen3.8-27B with LMCache recipe covering host-memory cache reuse, optional cache compression and filesystem offload. Packaging it into the image puts that cache offload path into the standard Docker release rather than a manual side install.

A release cadence you can count

The Releases.md page lists every Docker image tag with a month. Reading it: the vLLM line starts at 0.2.0-b2 in July 2025 when the repository was weeks old, and the pattern since has been roughly one release or more every month through September 2026. The only month in that window with no listed vLLM or Omni release is February 2026.

The May 2026 entry matters to the hardware story: per the README, B70 support arrived as a new platform image with the 1.4 tag, so the bigger card shipped with the same serving pipeline the B60 had already been running for months.

What the September vLLM images changed

Per the README changelog, the 0.26.0-b1 image upgraded the underlying vLLM to 0.26.0, improved time-to-first-token for Qwen models, and delivered what Intel lists as a series of quantization, LoRA serving and FP8 KV cache optimizations and fixes. The follow-up 0.26.0-b2 image fixed prefix caching with multi-token prediction, corrected the sym_int4 output path, improved MTP decoding performance, and fixed block-FP8 errors that appeared during concurrent decoding.

That read as two releases in one month, from one Intel engineering team, both carrying fixes for failure modes other than feature adds. The vLLM feature list in the README covers INT4 and FP8 quantized online serving plus pre-quantized FP8 models, embedding and reranker support, multi-modal and Omni model support, tensor, pipeline and data parallelism, maximum context length discovery, and a multi-modal web UI.

• Quantization: online INT4 and FP8 serving, plus pre-quantized FP8 and MXFP4 models in the supported models table.
• Parallelism: tensor, pipeline and data parallel, per the vLLM section of the README.
• Scope: text generation is one line, and the Omni image covers the image, video and audio side.

The Omni image: image, video and audio

Per the README, LLM Scaler Omni runs image, voice and video generation in two modes: an Omni Studio mode built on ComfyUI, and an Omni Serving mode delivering OpenAI-compatible API endpoints through SGLang Diffusion or Xinference. The ComfyUI side covers image generation and editing, text-to-video and image-to-video, video animation, 3D generation, text-to-speech and voice cloning, plus video upscaling. The Wan 2.2 T2V 14B flow gets an explicit multi-B60 demo in the README, and LTX-2 is listed among supported video models.

The Omni Serving side exposes an OpenAI-compatible images generation endpoint, a speech endpoint, and a transcription endpoint per the README, which keeps clients written for commodity APIs working unchanged against the B60 and B70. The August 0.2.0-b1 release, per the changelog, bumped the ComfyUI stack to 0.31 on XPU, added MiniMax H3 local video generation and Wan Animate 2 workflows, optimized Wan 2.2 14B T2V Turbo and LTX-2, and added quantized ComfyUI workflow support.

One caveat with a number on it: the README labels Omni experimental, per its own section heading. The commits kept moving on it anyway, with an October 9 commit upgrading ComfyUI to 0.39.0 on the XPU stack and aligning Torch 2.14 XPU integration, per the commit history.

The agent-skills layer for the same cards

A companion repository from Intel itself covers the same Arc Pro line from the agent-facing side: intel/gpu-ai-skills, a collection of Agent Skills for setting up, running, benchmarking and profiling Hugging Face safetensors and GGUF models on Intel GPUs including Arc, Arc Pro and Battlemage, per its README. The skills nominally cover the PyTorch plus Transformers, vLLM-XPU, SGLang-XPU and llama.cpp SYCL stacks, plus CUDA to XPU migration, so the sizing and serving knowledge about these cards is being written in a form an AI coding agent can consume directly. A sizing skill called model-can-it-fit estimates VRAM from the weights, KV cache and activations in a model's config.json, and its own example prompt is whether Qwen2.5-32B in int4 fits on an Arc Pro B70 at 8K context. The repository counts 40 stars, was pushed October 16, 2026, and is Apache-2.0 licensed, per the GitHub API.

The pairing matters more than either repo alone: LLM Scaler is the maintained serving stack, and gpu-ai-skills is the operating manual written for agents to run cards like the B70 themselves. A reader who covers one without the other sees half the software surface Intel is publishing for this hardware.

FAQ

What is Intel's LLM Scaler?
Per the official README, it is a GenAI serving stack for text, image and video generation on Arc Pro B60 and B70 GPUs, built on standard frameworks like vLLM, ComfyUI, SGLang Diffusion and Xinference and delivered as tuned Docker images.

Which GPUs does it target?
The README names exactly two: Arc Pro B60 and Arc Pro B70, the data-center-class Arc Pro line. B70 support arrived in the May 2026 platform image per the changelog.

What changed this week?
SGLang got explicit cache-consistent GDN, Q8 and fused MoE APIs on October 17; Omni was bumped to 0.2.0-b2 on October 16; and LMCache KVWeave cache offload was packaged for vLLM on October 14, all per the commit history.

How often is it updated?
Roughly monthly Docker image releases since August 2025, with February 2026 the only month in that window with no listed release, per the Releases page. The repository was pushed October 17, 2026.

Does it cost anything?
The code is Apache-2.0 licensed per the repository, and the framework layers are the standard open vLLM, SGLang, ComfyUI and Xinference projects. No purchase is involved; the deliverable is Docker images on Intel's registry.

Do the frameworks stay standard?
Per the README they are the same vLLM, ComfyUI, SGLang Diffusion and Xinference projects, tuned for the B60 and B70, so clients and workflows written against them do not become Intel-proprietary.

Our take

We covered the B70 board itself last week, and the part that decides whether that silicon earns its price is exactly this repo. A card nobody ports frameworks to is paperweight silicon, and Intel shipping cryptographically signed commits and timed releases monthly says someone is staffing the software story seriously. The untested caveat stands: we have not run the images, so we are reporting the changelog, not benchmarking it.

Bottom Line

Intel's LLM Scaler is the software layer that turns the Arc Pro B60 and B70 from boards into servable GenAI hardware, and the repository is actively maintained with framework-native vLLM, SGLang, ComfyUI and Xinference images releasing roughly monthly. This week the SGLang API surface, the Omni version and the vLLM cache-offload path all moved, per the commit history.

Sources: Per the official repositories intel/llm-scaler (README, Releases.md and commit history) and intel/gpu-ai-skills (README and GitHub API metadata), all fetched October 18, 2026.

Related: Intel GPU AI Skills Brings 20 Agent Skills for Hugging Face Models on Arc GPUs [2026] | Mach X Runs LLMs on Arc Pro B70 Cards: 16 Agents at Once [2026] | GUNNIR Packs Two Arc Pro B70 Cards Into One 64GB Workstation [2026]