Meta Muse Spark and the End of the Llama Era

**Meta Muse Spark** is Meta's flagship LLM released April 8, 2026 — the first Meta model since 2023 not released as open weights. ## Specs (disclosed) - **Artificial Analysis Intelligence Index (AAII)**: 52 - **Ranking**: #4 overall behind Gemini 3.1 Pro, GPT-5.4, Claude Opus 4.6 - **MMMU-Pro vision**: 80.5% (#2 behind Gemini 3.1 Pro Preview at 82.4%) - Natively multimodal with tool use, visual chain of thought, multi-agent orchestration - ~10x compute-efficient vs Llama 4 Maverick for equivalent capability - **Closed source** — no weights release - Parameter count undisclosed (rumored ~500B MoE) ## Deployment Muse Spark powers: - Meta.AI assistant - Facebook/Instagram/WhatsApp chat integration - Meta smart glasses (Ray-Ban Meta + Orion prototypes) - Internal Meta product teams ## Strategic shift Meta's open-source positioning was a defining strategic identity from Llama 1 (February 2023) through Llama 4. Key contributions to the open ecosystem: - Llama 1 (Feb 2023): leaked weights → ignited open-weight LLM movement. - Llama 2 (July 2023): official open-weights release, permissive license. - Llama 3 (April 2024): further performance jump, solid open-weight baseline. - Llama 4 (March 2025): ambitious MoE, positioned as catch-up to frontier closed models, disappointing reception. The Llama era made Meta the anchor of Western open-weight AI. Most open-source tooling (llama.cpp, ollama, lm studio, many fine-tuning libraries) was built around Llama ecosystem expectations. Muse Spark's closed status is the end of that era. Meta's framing: they need commercial monetization of AI, can't justify giving away frontier capability. Contrast with GLM 5.1 Open-Weight Model released one day earlier, which was beating closed-source models on real benchmarks while still MIT-licensed. ## Effect on the open ecosystem - The anchor US open-weight lab just left. - Remaining Western contributors: Mistral (partially moved to commercial-restricted), some research labs. - Open-weight ecosystem is now dominated by Chinese labs: Alibaba Qwen, DeepSeek, ZAI GLM, MiniMax. - Sovereignty concern: Western democracies deploying Chinese-origin open-weight models raises supply-chain questions similar to those about TikTok and Huawei networking gear. ## Reception Muse Spark itself is technically competent — solid #4 on AAII, big vision jump, efficient. Reception was mixed: engineers respected the capability, open-source community felt betrayed, enterprise buyers were neutral-to-positive (they prefer closed models with SLA). ## Meta's broader AI strategy Meta's 2025-2026 AI reorganisation included: - Acquiring Scale AI's data annotation business (2024). - Building the Meta AI Research super-group. - Massive H100/B100 infrastructure buildout (~600K+ H100-equivalent by end 2026, per disclosed estimates). - Shift to agentic and embodied research (smart glasses, Orion AR, Quest XR). Muse Spark is the model that powers this stack. The closed-source choice follows from treating it as a defensible competitive asset rather than an ecosystem contribution. ## Contrast with parallel week In the single week of April 7-12, 2026: - **ZAI** released GLM 5.1 Open-Weight Model at SOTA performance under MIT license. - **Meta** released Muse Spark closed. - **Anthropic** kept Claude Mythos Reward Hacking Behaviors closed with 'too dangerous' framing. - **MiniMax** released M2.7 with a tightened commercial-restricted license. The direction of the Western frontier is inexorably closed. The direction of the Chinese frontier is mostly still open but tightening. See AI News Week of April 12 2026 — Four Headline Stories.