Meta Muse Glimmer
Meta's 30B dense multimodal agent model. Apache 2.0, 131K context, 4-bit checkpoints under 20GB and none of the Llama licence conditions attached.
Overview
Muse Glimmer is a 30B dense multimodal model released 10 August 2026 by Meta Superintelligence Labs, built specifically for agents running on local hardware. It reports 27.85B active parameters at inference, a configured maximum context of 131,072 tokens and support for more than 100 languages. A dedicated perception encoder takes interleaved text and images so screenshots, charts and documents feed straight in. Reasoning effort is selectable at four levels from low to xhigh. Meta shipped a DFlash speculative decoding drafter with it, worth 3.1x on an RTX 5090 and 1.8x on an M5 Max.
The licence matters more than the architecture here. Llama 4, the last Llama, shipped in April 2025 under the Llama 4 Community License and Acceptable Use Policy: source-available rather than open source, with a requirement that anyone whose products had more than 700 million monthly active users in the month before the release date request a separate licence from Meta, plus a carve-out restricting multimodal use for entities domiciled in the EU. Glimmer drops all of it for Apache 2.0. Meta published full-precision and two 4-bit checkpoints under those terms, and it is the first Meta open release without the Llama conditions attached.
Practically it is aimed at the workloads that make sense on one machine. Function calling, long multi-turn tool sessions, local coding and using a model as a judge in an evaluation harness. The K-Quant-17GB checkpoint targets 24GB cards and K-Quant-Dynamic targets 32GB, with day one support across Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM and SGLang. The caution is institutional rather than technical. Meta's frontier model, Muse Spark 1.2 from 5 August 2026, remains closed, so treat Glimmer as one open release from a company whose position on this has moved twice in five months, not as a restored commitment.
Key Features
- ✓ 30B dense with 27.85B active at inference and a configured maximum context of 131,072 tokens
- ✓ Perception encoder takes interleaved text and images, so screenshots, charts and scanned documents go in directly
- ✓ Controllable reasoning effort at four levels from low through xhigh, so you can trade latency against depth per request
- ✓ DFlash speculative decoding drafter shipped with the weights, giving 3.1x speedup on an RTX 5090 and 1.8x on an Apple M5 Max
- ✓ Full precision plus two 4-bit checkpoints, the smaller one landing under 20GB
- ✓ Support in Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM and SGLang from release day, with over 100 languages covered
Where it holds
- • Apache 2.0 with none of the Llama Community License conditions: no 700 million monthly active user threshold, no acceptable use policy and no EU carve-out on multimodal use
- • Benchmarks aimed at the right targets for agent work, with SWE-Bench Verified at 76% and MCP Atlas at 75.5%, at a size that quantises under 20GB
- • 551,258 Hugging Face downloads in the first month, so quants, adapters and serving configs already exist rather than needing to be built
- • Built for what people actually run locally in 2026: function calling, long tool-use sessions, local coding and acting as a judge model, with an offline path
Where it breaks
- • Released 10 August 2026. Eighteen days of public use is not a track record, and every benchmark figure is vendor-reported
- • One model, not a family. No 7B for edge, nothing above 30B and no counterpart to the sizes Llama used to cover
- • Meta's frontier line stays closed. Muse Spark 1.2 shipped 5 August 2026 with closed weights, and while Zuckerberg said the weights would be opened, nothing had been published as of late August
- • Weights only. No training corpus, no training code and no statement about releasing either, so provenance questions have no answer
My Take
Some context is needed before the model itself. Meta ended the Llama line after Llama 4 in April 2025 and replaced it in April 2026 with the closed Muse Spark, which most people read as the end of Meta's open-weights position. Glimmer walks part of that back, and the licence change is substantive: Apache 2.0 means the Llama Community License is gone, including the clause requiring anyone above 700 million monthly active users to request a separate licence from Meta and the carve-out restricting multimodal use for EU-domiciled entities. Judge the model on eighteen days of evidence though. SWE-Bench Verified at 76% is a vendor number, half a million downloads is early enthusiasm rather than production validation, and one 30B checkpoint does not rebuild a family.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Apache 2.0
- Starting at:
- Free. Apache 2.0 weights on Hugging Face in full precision plus two 4-bit checkpoints, K-Quant-17GB aimed at 24GB cards and K-Quant-Dynamic aimed at 32GB. No first-party hosted API from Meta, so cost is hardware: a 24GB consumer GPU such as an RTX 4090 or 5090 runs the 17GB quant, an M4 or M5 Max Mac runs it on unified memory. Third-party endpoints exist but the model is designed to be run locally.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Meta Muse Glimmer holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Meta Muse Glimmer
Google Gemma 4
open-sourceGoogle's open-weight family in five sizes, 2B effective up to 31B dense, text and image in with audio on the smaller ones, 256K context and Apache 2.0 weights since April 2026.
Mistral AI
open-sourceEuropean lab publishing open weight models you can run on your own hardware. Large 3 is a 675B MoE with 41B active, 256K context and Apache 2.0 weights.
NVIDIA Nemotron 3
open-sourceNVIDIA's open-weight family, 31.6B Nano up to the 550B Ultra MoE. Hybrid Mamba-Transformer, 1M context, weights plus roughly 3T tokens of training data published.
InternVL 3.5
open-sourceShanghai AI Lab's open vision-language family (书生·万象), nine sizes from 1B to 241B-A28B under Apache 2.0. The default open VLM for OCR, document parsing and GUI agent work.