Meta's Muse Spark Open: 816B MoE coding model with 1M context, tuned on NVIDIA GB300 NVL72
sergeykarayev · x · 2026-11-18
Meta launched Muse Spark Open, an MoE model built for agentic coding: 816B total parameters, 42B active per token via top-8 routing across 256 experts, trained on NVIDIA GB300 NVL72 Blackwell Ultra.
Key specs:
- 1M token context: holds an entire repo — complex dependencies, legacy code, thousands of files — without task splitting.
- Multimodal perception: adds audio and video understanding alongside image/document input via a 1.8B perception encoder.
- Throughput: up to 50K tokens/sec per GPU and 600 tokens/sec per user on GB300 NVL72.
- NVFP4 checkpoints: shrink memory for all 816B parameters with native FP4 Tensor Core acceleration.
Engineering path: post-train on domain data with NVIDIA NeMo AutoModel (PyTorch-native, Hugging Face checkpoints, no conversion) supporting full SFT and LoRA, or deploy via a one-command NVIDIA NIM. Weights are on Hugging Face.
Notable as a pattern: frontier model releases now ship with a full fine-tuning and inference stack attached.
More from coding & agent
- How one developer built a robot companion in 24 hours with Astra — gabrielchua · 2026-11-18
- Meta launches Muse Spark Open: 816B MoE coding model with 42B active params and 1M token context — PyTorch · 2026-11-18
- Vercel launches eve 1.0, an open-source framework claiming production agents as easy as using them — soleio · 2026-11-18
- SC26 BoF: top HPC names to debate coding agents' threat to numerical linear algebra trust — kfountou · 2026-11-18
- Microsoft Copilot launches Workspaces, a shared-context space for teams and AI agents — testingcatalog · 2026-11-18
- Raindrop ships MCP analytics: traces, issues, intent clustering for MCP servers — ycombinator · 2026-11-18