AI Model Timeline

Tracking the accelerating release frequency of frontier AI models.

Muse Glimmer

Meta·

LLMsopen-weightEdge / Consumer

Overview

30B dense agentic model distilled from Muse Spark 1.2 and released under Apache 2.0 — Meta's first open-weight release since the Llama 4 family. At 4-bit the checkpoint stays under 20 GB, so the full setup including KV cache and perception encoder fits a 24–32 GB memory envelope on a single consumer GPU or Mac. A separate perception encoder handles image input; block-level speculative decoding keeps latency inside a real agent loop.

Capabilities and innovations

30B dense parametersMultimodal input via separate perception encoderAgentic tool calling with planning and failure recoveryControllable reasoning effort levelsTrained on 100+ languagesDistilled from Muse Spark 1.2Block-level speculative decoding for agent-loop latency4-bit K-Quant checkpoint under 20 GBInterleaved text/image via a dedicated perception encoder

Benchmarks

BenchmarkScoreSource
GPQA Diamond83.5%vendor
GPQA Diamond 83.5 from Meta's launch comparison table (Muse Glimmer vs Gemma4-31B vs Qwen3.6-27B). The table is embedded as an image; value read from press transcriptions of that table, not independently reproduced. Meta states it used the more favourable of a competitor's self-reported score or its own reproduction, and Artificial Analysis where all three models were covered.
Humanity’s Last Exam (no tools)21.96%third party
HLE 21.96% no-tools, measured by Artificial Analysis at high effort (accessed 2026-08-17). Replaces the evaluation_pending placeholder: Meta's own HLE figure is only available inside an image table whose methodology report was unreachable.
SWE-bench Verified76%vendor
SWE-bench Verified 76.0 from the same launch table. Era-1 benchmark, recorded only because Meta reports it.
SWE-bench Pro51.2%vendor
SWE-bench Pro 51.2 from the same launch table.
Terminal-Bench 2.x51.7%vendor
Terminal-Bench 2.1 51.7 from the same launch table. No verified entry on the official Terminal-Bench leaderboard as of 2026-08-10.
MMMU-Pro74%vendor
MMMU-Pro 74 from the same launch table.
  • MMLU-Pro: independent evaluation pending — Not identifiable in Meta's launch table (image); no independent MMLU-Pro run for Muse Glimmer published as of 2026-08-10.

Architecture and hardware

Parameters
30B total (Dense)
Estimated VRAM at Q4
~18 GB, Edge / Consumer class
Quantization formats
4-bit K-Quant, GGUF, MLX, BF16
Recommended runtime
Ollama
License
Apache 2.0

Links

More from Meta

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.