Meta Revives Open‑Source AI: 30B‑Parameter Muse Glimmer Runs on a Single GPU

Meta released the weights of Muse Glimmer, its first open‑weight model since Llama 4, under an Apache 2.0 license, allowing free deployment, fine‑tuning, and creation of derivative models.

Muse Glimmer contains roughly 29.6 B parameters, including a 1.8 B visual encoder, supports both text and image inputs, and provides a 128 K context window. It is positioned as the base for personal‑device agents, with emphasis on multi‑step reasoning, tool invocation, and fault recovery.

The full‑precision BF16 checkpoint is about 60 GB, which exceeds the capacity of typical desktops. Meta therefore offers a 4‑bit quantized version under 20 GB, enabling execution on machines with 24 GB–32 GB GPU memory such as high‑end PCs equipped with RTX 5090 or high‑spec Macs.

According to Artificial Analysis, Muse Glimmer scores an IQ of 35, 21 points higher than Llama 4 Maverick, close to Kimi K2.5 (36), and slightly below Qwen3.6 27B and Ling 3.0 Flash (both 38). In Meta’s own agent evaluation, Muse Glimmer achieves 75.5 on MCP Atlas (vs. Qwen3.6 27B’s 62.5), 74.6 on DeepSearch QA (vs. 71.1), and leads the τ3‑Banking tool‑use test with 23.5 points.

Independent tests expose weaknesses: on GDPval‑AA v2 the model reaches 953 Elo, below the human baseline of 1000 and behind Qwen3.6 27B’s 1141; its hallucination rate on AA‑Omniscience is 82 % versus 49 % for Qwen3.6; and it scores 52 % on Terminal‑Bench 2.1, again lower than Qwen3.6’s 61 %.

Mark Zuckerberg argues that open AI should be available to billions of users to balance institutional power, rejects safety‑risk arguments that justify keeping the strongest models closed, and proposes an independent board to approve safety standards while offering checkpoint data to governments for early risk detection.

The move also reflects market pressure: Chinese open models such as DeepSeek, Kimi, and Qwen are closing the performance gap with U.S. closed‑source leaders. By lowering the barrier for local deployment, Meta hopes to win back developers, while still monetizing inference, hosting, and tooling services. The company plans to invest up to $145 billion in AI infrastructure and has announced a $1 billion data‑center community fund.

If Muse Spark 1.2 is opened as promised, the competition between open‑source and closed‑source models will intensify, forcing the industry to decide which capabilities to open, who bears the cost, and how governance rules are set.