vLLM Ships Day-0 Support for Meta's 816B Open-Weight Muse Spark Open
woosuk_k · x · 2026-10-16
vLLM announced Day-0 support for Meta's open-weight Muse Spark Open, verified on NVIDIA, AMD GPUs and TPUs: 816B total params with 42B active per token, 1M-token context, native image/video/audio inputs, built for long-horizon agentic coding. Highlights: BF16/NVFP4/MXFP4 quantization and DSpark speculative decoding on both CUDA and ROCm. vLLM calls it the most capable US open-weight model, built with Meta Superintelligence Labs.
More from coding & agent
- Dev uses Claude to draw animated mascot GIFs frame-by-frame in code until usage runs out — jeremy_goldberg · 2026-10-16
- Bash's hidden /dev/tcp stack lets you make raw TCP calls without curl or netcat — RealGeneKim · 2026-10-16
- dair-ai explores context graphs for research tasks: LLM agents love structured data — omarsar0 · 2026-10-16
- Codex CLI v0.163.0 adds mid-response steering, parallel read-only tools and /config search — github-actions[bot] · 2026-10-16
- Hooking Codex CLI into Claude spawns a four-subagent "council" for any question — RosieCampbell · 2026-10-16
- AI-optimized MACS3 merged: 41-64x callpeak speedup via Claude Code, human-reviewed — jmschreiber91 · 2026-10-16