Meta today released Muse Glimmer, a roughly 29.6 billion-parameter dense multimodal transformer designed to run autonomous AI agents locally on consumer-class hardware. The release is notable both for the model’s agent-focused design and for its license: Meta published Glimmer’s weights under the permissive Apache 2.0 license.
Availability and rollout
- The model weights are available on Hugging Face as of the release. Meta says runtime support will roll out this week across Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter, with optimized integrations for llama.cpp, MLX and ExecuTorch arriving in the coming days. Meta names Unsloth as a local-runtime partner and points to PyTorch’s TorchTitan for fine-tuning.
- Mark Zuckerberg, Meta co-founder and CEO, announced the weights on X (under his longtime handle @finkd) and wrote that Muse Spark 1.2 weights will be released soon as well.
Technical spec highlights
- Parameters: ~29.6B total parameters across 52 layers, including a dedicated ~1.8B-parameter ViT-G/14 perception encoder.
- Inputs/outputs: accepts interleaved text and images; generates text. Supports 100+ languages.
- Context length: 131,072 tokens or more.
- Knowledge cutoff: January 4, 2026.
Meta trained Glimmer specifically around the agent loop—planning, calling tools, interpreting results, continuing work and recovering from failures—rather than primarily as a general chatbot. Alexandr Wang, Meta’s chief AI officer, said Glimmer “can operate as a fully capable agent” and that it “can run on 24GB of VRAM without losing agentic reliability.”
Training and fine-tuning
According to Meta’s technical description, Glimmer was pre-trained on Muse Spark outputs with logit distillation, mid-trained on long-context, agent-heavy datasets with richer reasoning traces, and post-trained with supervised fine-tuning, on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains.
Meta demonstrated the model with a local Home Assistant workflow in which Glimmer autonomously discovers a Home Assistant instance, queries device APIs, writes a responsive HTML/CSS/JavaScript dashboard, deploys a local server and verifies the result—illustrating an agent that must maintain a plan while interacting with external systems.
Hardware, quantization and memory targets
- At full precision, Meta says the model requires over 55 GB of memory, exceeding single consumer GPUs. To enable local deployment, Meta provides ~4-bit quantized builds that shrink language-model weights to under 20 GB and leave room for KV cache, the perception encoder and a companion speculative-decoding model.
- Two notable quantized targets: K-Quant-17GB (targeting 24 GB systems) and K-Quant-Dynamic (targeting 32 GB systems). These allow running the stack on high-end consumer GPUs such as Nvidia RTX 3090 or RTX 4090 (24 GB) and the newer RTX 5090 (32 GB). On Apple hardware, unified memory on M4/M5 Max MacBook Pros and Mac Studio with 32 GB+ can hold the stack; Meta ran its own speed tests on M4 Max and M5 Max.
- Typical 8 GB or 16 GB laptops remain insufficient. The full-precision BF16 release, pegged by Meta at ~64 GB, is intended for data-center GPUs or top-spec Mac Studio configurations.
Meta reports average accuracy degradation of about 0.2% across 15 benchmarks for the K-Quant-Dynamic (32 GB target) build, and about 1% for the K-Quant-17GB (24 GB target) build—figures from Meta’s own measurements, not independent evaluations.
Latency: DFlash speculative decoding
To tackle latency for agent workloads, Meta ships DFlash speculative decoding: a smaller drafter model proposes blocks of tokens (e.g., 16) that the primary model verifies in parallel. Meta’s reported speedups:
- Nvidia RTX 5090: 74.9 tokens/s → 233.4 tokens/s (3.1×)
- Apple M5 Max: 26.6 tokens/s → 50.2 tokens/s (1.8×)
- Apple M4 Max: 23.7 tokens/s → 37.8 tokens/s (1.5×)
Tests used batch size 1 and greedy decoding; Apple systems were measured via ExecuTorch and the RTX 5090 via llama.cpp. For agents, these multipliers matter because single user requests can trigger many model turns, tool calls and verification steps, and latency accumulates across them.
Benchmarks and market position
Glimmer competes in a field that includes Google DeepMind’s Gemma 4 family and Alibaba’s Qwen3.6-27B. Meta’s internal comparisons show Glimmer leading on several agentic tests (e.g., MCP Atlas 75.5; DeepSearch QA 74.6; τ³-Banking 23.5; WildClawBench 47.6; GAIA2 43.3) and scoring 51.2 on SWE-Bench Pro (versus Gemma4-31B at 36.9 and Qwen3.6-27B at 50.2).
However, Glimmer does not dominate all benchmarks: Qwen outperforms Glimmer on OSWorld-Verified (75.6 vs. 65.9), TerminalBench 2.1 (60.7 vs. 51.7), SkillsBench and several multimodal evaluations. Read together, these results suggest Glimmer is of particular interest as a specialized local-agent model rather than as an undisputed overall leader.
Safety assessment
Local agents change the threat model versus chat-only models. Meta’s safety metrics show mixed results: on the CI Memories privacy benchmark lower is better and Glimmer scores 26.4 (Gemma 12.1; Qwen 53.4). On the Siren AgentDojo prompt-injection test, Glimmer has a 28.4% attack-success rate (Gemma 25.6%; Qwen 40.3%) while achieving the highest utility score among the three at 94.2.
Meta evaluated Glimmer under its Advanced AI Scaling Framework and concluded it does not meet the framework’s definition of “Frontier AI,” because it is generally less capable than Muse Spark. The Preparedness Team assessed Glimmer at Moderate or lower risk across chemical/biological, cyber and loss-of-control categories. Meta nevertheless recommends deploying Glimmer inside broader systems with guardrails (for example, human-in-the-loop confirmation for irreversible actions).
Licensing, ecosystem and implications
- Meta is publishing full-precision BF16 weights, both 4-bit quantizations, the DFlash drafter and the perception encoder under Apache 2.0. Apache 2.0 permits unrestricted commercial use, modification and redistribution.
- The release makes the developer workstation a credible deployment target for autonomous agents rather than only an experimental environment. By combining a 30B dense model with quantized stacks tuned for 24–32 GB hardware and a permissive license, Meta gives developers broad freedom to run, modify and integrate a locally-hosted agent stack.
- Meta has not released the model training data or training code; the open element is the published weights.
What to watch next
The important test will be whether Glimmer’s benchmark strengths translate into robust, long-running real-world agent workflows that interact with repositories, enterprise tools and third-party systems. If so, a class of agents that previously required cloud APIs could increasingly run and act directly on developers’ local machines—shifting where sensitive context is processed, who pays for inference latency and how agentic applications are architected.
