Meta's 30B Apache-2.0 Agent Model Moves the Judge Workload Off the Invoice

Meta AI Research has released Muse Glimmer, a 30-billion-parameter open-weight model under the Apache 2.0 licence, built for always-on local agent work rather than conversation ^[1]. According to InfoQ's report, the pitch is that autonomous agents, complex tool invocation, local coding and LLM-as-a-judge evaluation can now run on consumer GPUs and workstations without a cloud API in the loop ^[2].

The part an operator can verify without trusting anyone is the memory budget. An uncompressed 30B model typically needs more than 55 GB of VRAM, which puts it outside consumer hardware ^[3]. Muse Glimmer uses 4-bit dynamic compression (K-Quant) to land at roughly 17 GB to 20 GB ^[4], a reduction of about 2.75x to 3.2x ^[5]. Meta says that fits inside a 24 GB to 32 GB GPU or NPU envelope with room for the KV cache, perception embeddings and speculative decoding overhead ^[6]. Run the subtraction and the range matters: at the top of the weight range on a 24 GB card you have about 4 GB left for everything else, while 17 GB weights on a 32 GB machine leave about 15 GB ^[7]. Extended agentic sessions grow the KV cache, and the model also carries a 1.8B-parameter perception encoder for reading screenshots, diagrams and documentation inline ^[8]. The recommended systems, an M4/M5 Max Mac or an RTX 5090 or 4090 ^[9], are effectively the honest floor rather than the entry point.

Throughput comes from DFlash speculative decoding, where a lightweight drafter proposes multi-token blocks that the base model validates in parallel, for up to a 3.1x gain on Apple Silicon M4/M5 Max and RTX 5090 hardware ^[10]. That is an upper bound on two named platforms, not a portable number.

The economic argument is strongest for the least glamorous workload. LLM-as-judge evaluation is high volume and repetitive, which is exactly the shape that per-token cloud pricing taxes hardest, and Apache 2.0 weights running locally remove both the meter and the data egress conversation ^[1][2]. Meta trained the model to handle failure states, diagnosing an errored API call or terminal command and trying alternative paths instead of stopping ^[11], and it supports adjustable reasoning effort and the OpenClaw agent framework ^[12].

What is missing is numbers. InfoQ reports evaluations on SWE-Bench, DeepSearch QA, tau-Bench and MCP-Atlas with "strong success rates" against leading open models in the 30B class, and claims superior multi-step tool reliability and failure recovery against Gemma 4 31B and Qwen 3.6 27B, but publishes no figures ^[13][14]. The accompanying architecture and benchmark figure is credited to Sebastian Raschka rather than to a Meta technical report ^[15]. Treat the tool-reliability claim as untested until someone runs it.

Distribution is the detail that decides whether this gets tried this week: weights are on Hugging Face with native execution through llama.cpp, ExecuTorch, Apple MLX, Ollama, LM Studio and vLLM, and fine-tuning through PyTorch's TorchTitan ^[16].

Watch for independent tau-Bench and MCP-Atlas runs against Gemma 4 31B and Qwen 3.6 27B ^[14], for whether the 3.1x drafter speedup survives on hardware Meta did not name ^[10], and for how the 24 GB configuration behaves once a long tool-calling session has filled its cache ^[6][7].