Meta Superintelligence Labs released Muse Glimmer 30B on August 10, 2026—the first open-weights model to emerge from the division since the $14.3 billion Scale AI acquihire in June 2025. The release is a strategic pivot: Meta is betting that the future of agentic AI belongs not to the cheapest API but to the most accessible local deployment.
The model is a 29.6B dense causal transformer paired with a 1.8B ViT-G/14 perception encoder, released under the Apache 2.0 license. By providing the weights, quantizations, drafter, and perception encoder as open assets, Meta is positioning Glimmer as the foundational layer for developers building agent-native applications that require low-latency, private, and persistent execution on consumer hardware.
The architecture is tuned for that purpose. Glimmer fits within 24–32GB of VRAM—approximately 20GB at 4-bit quantization—placing it within reach of a single high-end consumer GPU. To address the latency bottleneck that typically makes local inference impractical for real-time agents, Meta introduced DFlash speculative decoding. DFlash proposes blocks of 16 tokens that the main model verifies in parallel, preserving output quality while delivering substantial throughput gains. On an Nvidia RTX 5090, DFlash accelerates generation from 74.9 to 233.4 tokens per second—a 3.1x speedup. On Apple’s M5 Max, the gain is 1.9x; on the M4 Max, 1.6x.
That hardware story matters because it separates Glimmer from the current generation of frontier open-weight models. Kimi K3 from Moonshot AI and DeepSeek V4 Pro are trillion-scale Mixture-of-Experts architectures with 2.8T and 1.6T total parameters, respectively. Both deliver strong benchmarks, but their size makes them impractical for most local deployments. Glimmer’s 30B dense architecture is a deliberate trade: fewer absolute parameters for hardware accessibility, enabling always-on local execution that larger MoE models cannot match on consumer systems.
The benchmark picture reflects that trade. In high-reasoning mode, Glimmer scored 51.2 on SWE-Bench Pro—the contamination-resistant agentic coding benchmark—leading Gemma4-31B (36.9) and Qwen3.6-27B (50.2). On SWE-Bench Verified, it scored 76.0, trailing Qwen3.6-27B by 1.2 points. On MCP Atlas Public, which tests tool-calling and multi-step agent workflows, Glimmer reached 75.5, outperforming both comparators by a wide margin.
Meta has not published a hosted API price for Glimmer. Instead, the model is available on Hugging Face for local inference and via third-party providers. Together AI lists it at $0.35 per million input tokens and $1.50 per million output tokens—well below Kimi K3’s $3/$15 but above DeepSeek V4 Flash’s $0.09/$0.18. The pricing is beside the point. Meta is not building a revenue stream from API calls; it is building an ecosystem. The model supports seven runtimes at launch: llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, vLLM, and SGLang—the same runtimes that power the majority of existing local agent deployments.
The $14.3 billion Scale AI deal brought Alexandr Wang into Meta to lead MSL, reporting directly to Mark Zuckerberg. Scale AI retained its independence under interim CEO Jason Droege. Glimmer is the first open-weight model released under Wang’s leadership of the division, following the proprietary Muse Spark 1.2 and the closed-source Muse Code agent earlier this year. The choice to go open-weight on the first model optimized for local deployment—rather than continuing the closed-source approach—suggests a deliberate strategy to prioritize adoption over control.
The competitive dynamic this introduces is structural. Cloud-based model providers—including Meta’s own Meta Model API—will continue to offer frontier-scale reasoning. But for the emerging class of always-on local agents that handle scheduling, coding, file management, and multi-step tool use, the constraint is not raw capability. It is the ability to run persistently on hardware people already own. Muse Glimmer is Meta’s bid to define that category before anyone else does.