Meta AI Updates: November 11, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. A PyTorch Post Details Serving Meta’s 816B-Parameter Muse Spark Open on a Single TPU v7 Host With a Megakernel
Meta / PyTorch. A PyTorch blog post published November 10 describes serving Muse Spark Open, Meta’s open-weights mixture-of-experts model with 816B total parameters and about 42B active per token, on one TPU v7 (Ironwood) host of four chips using SGLang on the TorchTPU backend. A megakernel runs all 62 layers of a decode or verify step as one Pallas program; with DFlash2 speculative decoding, the post reports 1.95 ms per output token for a single request against 3.78 ms for the previous path, and a 1.85x gain for two concurrent requests. Other changes include a sequence-on-lane fp8 KV layout that halves KV bytes per token, host-memory L2 HiCache that raised throughput by up to 77% on agent traffic, and a fused MoE prefill kernel that cut a 4,096-token chunk from 440 to 384 ms. On SemiAnalysis’s 1M-context AgentX benchmark the post reports a curve from 433 output tokens per second per request at 6,101 tokens per second per chip to 25,906 tokens per second per chip at 95 tokens per second per request. The code is open source in the SGLang-TorchTPU repository on a development branch. Source
2. ExecuTorch Adds a WebGPU Backend to Run PyTorch Models in the Browser, 4x to 7x Faster Prefill Than Its Wasm CPU Path
Meta / PyTorch. The PyTorch team announced an ExecuTorch WebGPU backend on November 10. Models exported with torch.export to a .pte file run in WebGPU browsers without conversion to another framework format, and the same file also runs natively on Metal, Vulkan and D3D12. Compute runs through WGSL shaders and the C++ runtime compiles to WebAssembly. In Chrome on an Apple M1 Pro with quantized Llama 3.2 1B and Qwen3-0.6B, WebGPU was 4x to 7x faster than ExecuTorch’s WebAssembly XNNPACK CPU path on prefill and 2x to 10x faster on decode. Validated models include Llama 3.2 1B, Qwen2.5-0.5B, Qwen3-0.6B, Gemma 3 1B, SmolVLM, SAM 2, Depth Anything V2, Whisper base and Real-ESRGAN, plus LoRA and DiReFT fine-tuning for three small LLMs. The post says the backend is in early development, and a browser playground and Hugging Face model repos with export recipes are available. Source