Meta’s official open-weight release of Muse Spark Open, an 816 billion-parameter any-to-text model, represents a major milestone in multimodal AI and the most capable U.S. open-weights model released to date. Deploying an 816 billion-parameter architecture requires extreme memory efficiency and hardware-level low-precision acceleration. Intel has enabled day-0 support across Intel® Xeon® processors and Intel data center GPU platforms, Intel® GPU D-series, through targeted low-bit model quantization.
Muse Spark Open is a coding-optimized open-weight model. Trained specifically for agentic coding workflows, it delivers higher first-attempt accuracy, reliable tool calling and deep codebase understanding across the repositories developers use.
This open-weight model is strong in code generation, debugging and codebase understanding. The 1 million-token context window handles entire projects, including complex dependencies, legacy code and thousands of files, without requiring tasks to be broken into smaller pieces.
Muse Spark Open also supports multimodal understanding, including audio, video, image and document comprehension.
For agentic use cases, Muse Spark Open was built to plan, act, observe and self-correct over extended workflows. It is trained to browse the web, including multimodal content, use an agentic harness to spawn and coordinate subagents for decomposable work, and leverage computer-use capabilities across desktop, mobile and browser environments.
Data Center Silicon Built for 816 Billion-Parameter Scale
Large language models at the scale of Meta’s Muse Spark Open highlight the growing infrastructure demands of advanced AI workloads. At hundreds of billions of parameters, they require substantial memory capacity, efficient low-precision compute, power efficiency and architectures that can bring these capabilities into existing data center footprints.
Muse Spark Open has been tested and is running on the upcoming Intel® GPU D-series, today. This early work reflects Intel’s commitment to day-0 support for emerging AI models and gives customers an early path to begin evaluating Muse Spark Open on Intel silicon.
The image above shows Muse Spark Open inference running on four Intel® GPU D-series GPUs. Each GPU provides 160 GB of memory, offering sufficient capacity to run Muse Spark Open across four GPUs.
Intel® GPU D-series is designed for the demands of large and agentic AI models. Built on the Intel® Xe3P architecture, it features:
- 32 Xe3P cores
- 256 Intel® XMX engines
- Dedicated hardware acceleration across a broad range of data types, including MXFP4
- 160 GB LPDDR5x memory capacity, with options up to 480 GB
These capabilities are well aligned with large-model inference, combining low-precision compute with high memory capacity to support models such as Muse Spark Open efficiently. Its 350W air-cooled PCIe form factor also simplifies deployment. A compact four-card configuration can host an 816 billion-parameter-class model within conventional air-cooled data center racks, helping reduce infrastructure complexity for organizations evaluating frontier-scale AI.
Muse Spark Open on Intel® Xeon® Processors
Designed for high-capacity CPU inference, Intel® Xeon® processors support INT4 weight-only quantized (WOQ) inference, reducing memory footprint while leveraging Intel® Advanced Matrix Extensions (Intel® AMX) to accelerate AI matrix operations.
Given the size of the model, Intel® Xeon® processors provide a practical path for offline inference by taking advantage of large system memory capacities, enabling organizations to run workloads privately and replicate deployments across existing Intel® Xeon® infrastructure in their data centers.
Intel is announcing day-0 support for Muse Spark Open on Xeon processors. For step-by-step instructions on running Muse Spark Open on Xeon processors with SGLang, see the SGLang cookbook. Intel is also working closely with the vLLM community to add Xeon recipes.
Software Support and Enablement
Intel’s Upstreaming First software strategy helps enterprise AI engineering teams deploy both quantized variants using standard open-source tools.
Full upstream support for compressed tensors, such as MXFP4, and GPTQ (INT4 WOQ) in mainline vLLM enables hardware-accelerated matrix execution, continuous batching and PagedAttention across multinode deployments. Native upstream integration for microscaling (MX) execution and Intel XPU-optimized kernels in SGLang drives high-throughput continuous batching and low-bit inference for complex, multiturn agentic workflows.
To maximize efficiency on Intel silicon without sacrificing accuracy, Intel quantized MXFP4 and INT4 variants directly from Muse Spark Open’s original BF16 weights using vLLM’s native quantization tool, llm-compressor, powered by the AutoRound quantization algorithm.
Intel’s AutoRound algorithm optimizes rounding and clipping-range parameters layer by layer, helping preserve Muse Spark Open’s fine-grained reasoning capabilities even at ultra-low bit widths.
The MXFP4 weights, in compressed-tensors format, target Intel® GPU D-series GPUs to use native hardware acceleration, while the INT4 weights, in GPTQ format, support Xeon processors in weight-only quantization mode to minimize memory footprint and latency.
Combining upstream support in vLLM and SGLang with hardware-tailored MXFP4 and INT4 WOQ execution gives enterprise teams the performance and agility needed to run Meta’s 816 billion-parameter Muse Spark Open on day 0.
Summary
Intel® Xeon® processors offer an alternative approach for offline inference with Meta’s Muse Spark Open 816 billion-parameter model. Additionally, the upcoming Intel data center GPU, Intel® GPU D-series, is designed for the demands of large and agentic AI models and provides 160 GB of memory per GPU. A compact four-card configuration can host Muse Spark Open within conventional air-cooled data center racks.