Meta open-sourced Muse Glimmer on August 10, 2026, a 30-billion-parameter dense multimodal model distilled from Muse Spark and purpose-built for always-on local agent workflows. Released under Apache 2.0, the model targets systems with 24–32 GB of VRAM and runs on consumer hardware without cloud API dependency. It excels at end-to-end agentic task completion, reliable tool use, multi-step reasoning, and failure recovery across frameworks including OpenClaw and MCP.
Muse Glimmer fits memory constraints through two optimizations: 4-bit dynamic quantization compresses weights to 17–20 GB, leaving headroom for KV cache and perception embeddings; and DFlash speculative decoding, a lightweight drafter model that proposes multi-token blocks, delivers up to 3.1x generation speedup on NVIDIA RTX 5090 (1.8x on M5 Max, 1.5x on M4 Max). Benchmarks show strong agentic orchestration on MCP-Atlas (75.5), DeepSearch QA (74.6), and SWE-Bench Pro (51.2), with a trailing edge on terminal/computer-use work where Qwen3.6-27B leads (75.6 vs 65.9 on OSWorld).
Days after Muse Glimmer's release, Alibaba launched Qwen3.8-27B—also dense, also ~27B parameters, also Apache 2.0 (per Alibaba's commitment)—with direct competitive positioning. On Terminal Bench 2.1, which measures agentic terminal coding, Qwen3.8-27B scores 73.0 vs Muse Glimmer's 51.7, and also leads on SWE-Bench Pro (61.7), GPQA Diamond (89.2), and IFBench (79.5). But Qwen3.8-27B lags on some general reasoning tasks where Muse Glimmer excels.
For architects: the open local-agentic model tier is now fragmented by task. Muse Glimmer suits reasoning-heavy agent orchestration; Qwen3.8-27B targets terminal/repo-level coding and office automation. Both fit a 24 GB consumer GPU, both ship as Apache 2.0 open weights, and both reduce token billing and latency vs. cloud APIs. This tier is displacing smaller closed models and forcing frontier labs to justify cloud inference costs against persistent locally-run alternatives. The benchmark gap narrows monthly.