Meta Superintelligence Lab open-sources Muse Glimmer-30B

Meta Superintelligence Lab has pushed a new open-weight multimodal model to Hugging Face, and on paper it is aimed squarely at the agentic-workloads-on-your-own-hardware crowd. According to the model card on Hugging Face, Muse Glimmer-30B is a roughly 29.6B-parameter dense causal transformer paired with a ~1.8B ViT-G/14 perception encoder, released under Apache 2.0 with a 131,072-token context window, a 202,048-token vocabulary, and support for 100+ languages.

The framing on the card is unambiguously agentic. Meta lists end-to-end agentic task completion, reliable tool use with correct schemas, long-horizon reasoning, and failure recovery as headline capabilities, and it advertises compatibility with scaffolds like OpenClaw and Hermes Agent. Users can dial reasoning strength across low, medium, high, and xhigh in the system prompt. The self-reported benchmark table is stiff: 94.7 on AIME 2026, 83.5 on GPQA Diamond, 76.0 on SWE-Bench Verified, 51.2 on SWE-Bench Pro, 74 on MMMU Pro, and 75.5 on MCP Atlas.

The more interesting story for most builders is the deployment envelope. The card puts BF16 memory at 64GB, drops that to 32GB with K-Quant-Dynamic at a stated 0.2% degradation, and to 24GB at 1.0% degradation. With Meta's DFlash speculative decoding, an RTX 5090 goes from 74.9 to 233.4 tokens per second, a claimed 3.1x speedup, and an Apple M5 Max moves from 26.6 to 50.2 tok/s. If those numbers hold up outside Meta's own benches, a top-end consumer workstation becomes a viable home for a serious agent stack instead of a monthly API bill.

The card is thin on the things you would actually want to audit. Training data is described only as publicly available data, third-party data, and Meta products/services, with no dataset list or filtering description. Every benchmark score is Meta's own, with no independent replication yet. And the safety section rates chem-bio, cyber, and loss-of-control risks as 'moderate or lower' without publishing the evaluation methodology. The listing already shows 292,973 monthly downloads and 56 open discussions, and two experts in our tracked directory have circulated the release.

For anyone building agent products, the pitch is a locally-runnable, Apache-2.0 multimodal model that can plausibly do tool-heavy coding and browsing work on a single GPU or high-end Mac. Whether the benchmark numbers survive contact with outside evaluators is the question the next few weeks will settle.