Open Source · AUGUST 11, 2026
Meta ships Muse Glimmer: a 30B open-weight agent for a single consumer GPU
Meta Superintelligence Labs distilled Muse Spark 1.2 into a 30-billion-parameter Apache 2.0 model that runs on 24–32 GB of VRAM, and pledged to open the teacher's weights within weeks.
Meta Superintelligence Labs released Muse Glimmer on August 10, 2026, a 30-billion-parameter dense model under Apache 2.0 that fits on a single consumer GPU. It's Meta's most credible open-weight shipment in more than a year, and the framing is unmistakable: after a stretch in which Alibaba, DeepSeek, and Moonshot defined what open weights meant, a U.S. lab is trying to reenter the conversation on the terms Chinese labs set.
Glimmer is distilled from Muse Spark 1.2, the closed-weight teacher that MSL chief Alexandr Wang's team has been holding back. Spark 1.2 scores 54 on Artificial Analysis's Intelligence Index against a class median of 32, so the parent is genuinely frontier-adjacent. Meta says Spark's weights will follow within weeks. That's the real news; Glimmer is the trailer.
The specs are engineered for local inference. A 131,072-token context window, a 202,048-token vocabulary, a knowledge cutoff of January 4, 2026, and multimodal training across text, images, and more than 100 languages (per TechCrunch). Full precision exceeds 55 GB, but 4-bit quantization compresses it under 20 GB, landing squarely in the 24 or 32 GB VRAM envelope of a high-end desktop card or an Apple Silicon laptop.
Runtime coverage is the tell. Meta says llama.cpp, MLX, and ExecuTorch integrations land within days; Ars Technica reports Ollama, LM Studio, vLLM, and SGLang are live or imminent. A speculative-decoding scheme called DFlash delivers 3.1x throughput on an NVIDIA RTX 5090, 1.8x on an Apple M5 Max, and 1.5x on an M4 Max. The SGLang team clocked roughly 230 tokens per second on a single RTX 5090 with NVFP4 and DFlash enabled. Meta wants Glimmer running on hardware people already own, on tomorrow.
On benchmarks, Meta claims Glimmer beats Gemma4-31B and Qwen3.6-27B across DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, evaluated inside its Advanced AI Scaling Framework. Independent replication, as Ars Technica notes, is still pending. Vendor charts remain vendor charts.
The strategic frame arrived alongside the model. Mark Zuckerberg published a 6,500-word essay on personal intelligence, and MSL is positioning Muse Code, a terminal coding agent powered by Spark 1.2, as the commercial surface above the open base. The three-phase build Meta describes on its blog is designed to let developers fine-tune Glimmer while paying Meta for Spark-tier inference.
Read structurally, this is the Llama playbook rewritten for the DeepSeek era: give away the distilled model, keep the teacher on API, and hope the ecosystem consolidates around your weights before someone in Hangzhou ships something smaller and faster. The pledge to open Spark 1.2's weights in weeks is the commitment that matters. Everything else is positioning until it lands.
Sources
- https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
- https://www.bloomberg.com/news/articles/2026-08-10/meta-releases-muse-glimmer-ai-model-people-can-run-on-their-laptop
- https://techcrunch.com/2026/08/10/metas-new-glimmer-ai-model-offers-a-hint-at-zuckerbergs-personal-intelligence-vision/
- https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html
- https://arstechnica.com/ai/2026/08/with-new-open-models-meta-pitches-another-reboot-of-its-struggling-ai-strategy/