AMD adds day-0 support for Meta's Muse Spark Open

AMD said on Thursday it has added day-0 support for Meta's Muse Spark Open on Instinct MI350X and MI355X GPUs using ROCm software and vLLM, positioning the stack for a large open-weight model aimed at coding and long-running reasoning tasks.

Evidence and context

In the technical article published Oct. 15, 2026, AMD described Muse Spark Open as the first open-weight model in the Muse Spark family and said it is a natively multimodal Mixture-of-Experts model with 816 billion total parameters, 42 billion active per token and a 1 million-token context window. AMD said the model includes configurable reasoning and native tool calling, and that it is available on Hugging Face, Muse Code, the Meta Model API and OpenRouter.

AMD said the model can run on a single node of eight Instinct MI350X or MI355X GPUs, with 288 GB of HBM3E per GPU and 8 TB/s memory bandwidth, using tensor parallelism across eight GPUs. The company said the software stack is ROCm 7.2 with vLLM, along with the AMD AITER kernel library for attention, MoE and communication, and CUDA-graph capture for the decode step.

The article listed three supported checkpoints for vLLM: a BF16 open release, an MXFP4 AMD Quark export and an FP8 AMD Quark export using PTPC. AMD said the BF16 checkpoint can also be served with an opt-in mode that rounds routed experts to MXFP4 at load time, giving the MXFP4 memory and throughput profile without a separate checkpoint.

AMD also published performance figures from its tests. It said decode-step optimizations increased the MXFP4 checkpoint from 206 to 350 output tokens per second at four concurrent requests and from 719 to 1,180 at 32 concurrent requests. The company said expert parallelism added 7% to 20% throughput at 32 to 128 concurrent requests on MXFP4 and reduced time to first token by about 24%, reaching 2,197 tokens per second at 128. AMD said its tuned AITER fused-MoE table raised single-request decode throughput from 98 to 114 tokens per second on TP8, while DSpark speculative decoding reached 199 tokens per second in a single stream.

AMD said GSM8K accuracy stayed within a 94.4% to 95.5% range across its MXFP4 configurations. The company recommended different serving setups depending on load, including TP8 plus EP8 plus DSpark for up to about 64 concurrent requests and TP8 plus EP8 for higher loads. AMD also said the day-0 recipe is available in vLLM-recipe and that its ROCm vLLM Docker image is vllm/vllm-openai-rocm:muse-spark-open.

The article did not provide commercial terms for the support, and the reported throughput and latency figures were AMD's own test results on the configurations it described.