Meta introduces Muse Glimmer: 30B open-weight local agent model under Apache 2.0

Background and pain point

Local agents have been stuck between two options: frontier closed models are strong but cloud-bound and hard to keep private, while open models small enough for a PC are shaky at tool use and multi-step planning. Meta's 30B open-weight Muse Glimmer, released in August, targets that gap; on October 19 Meta published its technical report, detailing architecture, distillation and quantization for the first time.

Architecture and mechanism

Per the report, Muse Glimmer is a dense model of about 29.6B parameters: 52 layers, hidden size 6656, plus a ~1.8B ViT-G/14 vision encoder, a 202,048-token vocabulary and 128K+ context. Attention interleaves local and global layers with only 2 KV heads per layer to bound KV-cache growth. Under a 32 GB VRAM budget the team compared dense against 1:8 and 1:16 MoE designs and found dense 2 to 2.5x more compute-efficient. It is logit-distilled from Muse Spark, then trained with SFT, RL and on-policy distillation. BF16 weights are 55.7 GB; quantized language-model weights fit under 20 GB, and with a DFlash draft head it runs at practical speeds on a single RTX 5090.

Benchmarks

Meta's figures versus Gemma4-31B and Qwen3.6-27B (both thinking): MCPAtlas 75.5 vs 54.2 and 62.5, tau3-Banking 23.5 vs 15.1 and 16.7, SWE-Bench Pro 51.2 vs 36.9 and 50.2, AA-LCR 80.0 vs 68.3 and 73.3. It does not lead everywhere: SWE-Bench Verified 76.0 trails Qwen's 77.2, TerminalBench 2.1 is 51.7 vs 60.7, OSWorld-Verified 65.9 vs 75.6, and GPQA Diamond 83.5 is below both.

Takeaways for builders

Weights are Apache 2.0 and run locally on PCs or Macs with 24 or 32 GB. Tool calling, MCP and long-context memory tasks are its strengths; for heavy terminal work and GUI control Qwen3.6-27B still scores higher, so mix by task. The report also describes four reasoning-effort levels set in the system prompt to trade latency for quality.