Meta Muse Glimmer
by Meta (Meta Superintelligence Labs)
Apache 2.0 open-weight 30B model built to run always-on AI agents on a single consumer GPU
Muse Glimmer is Meta's 30B open-weight, Apache 2.0-licensed model purpose-built for always-on local AI agents. It handles multi-step tool use, long task trajectories and failure recovery, ships with an integrated vision encoder and 4-bit builds that fit a single 24GB consumer GPU, and is aimed at platform and engineering teams that want agent workloads running on their own hardware rather than a metered cloud API.
Muse Glimmer is a 29.6-billion-parameter dense transformer released by Meta Superintelligence Labs on 10 August 2026 under an Apache 2.0 licence — the most permissive terms Meta has applied to a model of this class, and a deliberate break from the restrictive Llama Community Licence that complicated commercial deployment. It is engineered for one job: running always-on agents locally. The architecture comprises 52 transformer layers, grouped-query attention with 32 query and 2 key-value heads, an interleaved attention pattern of three local layers (2,048-token sliding window) to one global layer, a 202,048-token vocabulary, a context window of 131,072-plus tokens, a knowledge cutoff of 4 January 2026, and an integrated ~1.8B-parameter ViT-G/14 vision encoder accepting up to 4,096 visual tokens per image, making it multimodal rather than text-only. Meta trained it by logit distillation from its proprietary Muse Spark model, then mid-trained on agent-heavy data with reasoning traces, followed by supervised fine-tuning, on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains; reasoning effort is adjustable through the system prompt across low, medium, high and xhigh modes. Two 4-bit quantised builds ship alongside the BF16 weights: K-Quant-17GB targets 24GB cards such as the RTX 3090 and 4090 at roughly 1% average accuracy degradation, and K-Quant-Dynamic targets 32GB systems such as the RTX 5090 and M4/M5 Max MacBook Pro at 0.2%. A bundled DFlash speculative-decoding drafter predicts 16 tokens per forward pass for a 3.1x throughput gain on an RTX 5090 (74.9 to 233.4 tokens per second). Weights are distributed on Hugging Face with day-one support across Ollama, LM Studio, llama.cpp, vLLM, SGLang, ExecuTorch and Apple MLX, plus hosted inference from Together AI, Fireworks AI and OpenRouter. Reported results include MCP Atlas 75.5, SWE-Bench Verified 76.0 and AIME 2026 94.7, positioned against Google Gemma 4 31B and Alibaba Qwen3.6-27B.
Platform and ML engineering leaders who need agent workloads to run on owned or edge hardware — because of data residency, unpredictable per-token cloud spend, or offline requirements — and who can staff the operational work an unsupported open-weight model implies.
Production-grade agentic reasoning and tool use running entirely on a single 24GB consumer GPU under a licence with no commercial restrictions, removing both per-token API cost and third-party data exposure.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Free, Usage-based
- Target Market
- CTOs, Heads of AI, Platform Engineers, ML Engineers, Enterprise Developers
- Deployment
- Open-source, Self-hosted, Edge-first, API-based
- Founded
- 2004
- Headquarters
- Menlo Park, United States
- Team Size
- 500+
Key Features
- ✓Apache 2.0 open weights
Permissive licence with no commercial usage restrictions, covering BF16 weights, quantised variants, the DFlash drafter and the perception encoder — a material change from Llama's community licence for enterprise legal review.
- ✓Single-GPU agent execution
K-Quant-17GB fits 24GB consumer cards at roughly 1% accuracy loss and K-Quant-Dynamic targets 32GB systems at 0.2%, so a full agent runs on hardware teams already own.
- ✓Integrated multimodal perception
A ~1.8B ViT-G/14 vision encoder accepting 4,096 visual tokens per image lets agents read screens, documents and UI state rather than text alone.
- ✓DFlash speculative decoding
A bundled drafter predicts 16 tokens per forward pass, delivering 3.1x throughput on an RTX 5090 and 1.8x on an M5 Max, which is what makes local always-on agents usable.
- ✓Adjustable reasoning modes
Low, medium, high and xhigh reasoning effort is selectable from the system prompt, letting one deployment trade latency against accuracy per task type.
- ✓131K+ token context with agent-trained recovery
Long context plus training on failure recovery means the model retries alternative execution paths when an API call or terminal command errors instead of terminating the trajectory.
- ✓Broad day-one runtime support
Ollama, LM Studio, llama.cpp, vLLM, SGLang, ExecuTorch, Apple MLX and PyTorch TorchTitan for fine-tuning, plus hosted endpoints on Together AI, Fireworks AI and OpenRouter.
Capabilities
Use Cases
- •Local coding agent
Developers run a SWE-Bench-capable coding agent on a workstation GPU, keeping proprietary source code off third-party inference endpoints entirely.
- •Offline and air-gapped operations agents
Defence, industrial and healthcare teams deploy agents that operate on-site with no outbound network dependency, satisfying strict data-residency rules.
- •High-volume LLM-as-a-judge evaluation
Evaluation and grading pipelines that would cost heavily on per-token APIs run locally at fixed hardware cost across millions of samples.
- •Screen-reading and document agents
The bundled vision encoder lets agents interpret dashboards, scanned forms and application UIs without a separate OCR or vision service in the pipeline.
- •Fine-tuned vertical agents
Teams post-train on domain data with TorchTitan and redistribute the derivative commercially, which the Apache 2.0 licence explicitly permits.
- •Cost-tiered inference routing
Enterprises route routine agent steps to a self-hosted Muse Glimmer and escalate only hard reasoning steps to a frontier cloud model.
Ideal For
Best For
- ✓Running always-on local or edge agents that need tool calls and failure recovery without a cloud round trip
- ✓Regulated or air-gapped environments where prompts and documents cannot leave owned infrastructure
- ✓Replacing high-volume, low-complexity cloud API calls where per-token billing dominates the AI budget
- ✓On-device coding assistants and terminal agents on developer workstations with 24-32GB of VRAM
- ✓LLM-as-a-judge evaluation pipelines run at scale where hosted inference would be prohibitively expensive
- ✓Teams that need to fine-tune and redistribute a derivative model under a permissive licence
Not Ideal For
- ✗Teams needing frontier-class ceiling performance — Meta classifies it as 'moderate or lower risk', not a frontier-tier model, and it trails Qwen3.6-27B on OSWorld-Verified computer-use and TerminalBench 2.1 terminal tasks
- ✗Organisations without GPU operations capability: there is no vendor SLA, no support contract and no managed uptime, so you own inference reliability end to end
- ✗Laptops and workstations with 8-16GB of memory, which remain unsupported even at 4-bit quantisation
- ✗Security-sensitive agent deployments without added guardrails — Meta's own card reports a 28.4% prompt-injection attack success rate on Siren AgentDojo and a 26.4 CI Memories privacy violation rate, both worse than Gemma 4
Integrations
Deployment
Market Analysis
Pros
- ✓Apache 2.0 is the most permissive licence Meta has released for a model of this size, eliminating a common procurement blocker
- ✓Runs a full agent loop on a single 24GB consumer GPU, converting variable per-token API spend into fixed hardware cost
- ✓Day-one support across the entire local-inference ecosystem plus three hosted providers, so evaluation does not require custom serving work
- ✓Strong reported agentic and reasoning scores — MCP Atlas 75.5, SWE-Bench Verified 76.0, AIME 2026 94.7
- ✓Bundled DFlash drafter delivers a measured 3.1x speedup on an RTX 5090, which is the difference between a demo and an always-on agent
Cons
- ✗Safety posture is weaker than its closest peer: Meta's own reporting shows a 26.4 CI Memories privacy violation rate against Gemma 4's 12.1, and a 28.4% prompt-injection attack success rate on Siren AgentDojo against Gemma's 25.6%, so guardrails and human confirmation for irreversible actions are mandatory rather than optional
- ✗Does not sweep its benchmark set — VentureBeat's comparison has it trailing Qwen3.6-27B on OSWorld-Verified computer use (65.9 vs 75.6) and TerminalBench 2.1 (51.7 vs 60.7), which are exactly the tasks a desktop agent performs
- ✗No vendor support, SLA or indemnity of any kind; open weights mean the operational burden and the liability both sit with the adopter
- ✗Hardware floor excludes most corporate laptops — 8-16GB machines are unsupported, so 'runs locally' means on a workstation or a well-specified Mac, not on standard issue fleet hardware
- ✗Knowledge cutoff of 4 January 2026 means retrieval or tool access is required for anything current
Pricing
Open weights (Apache 2.0)
$0
- ✓BF16 full-precision weights
- ✓K-Quant-17GB and K-Quant-Dynamic 4-bit builds
- ✓DFlash speculative-decoding drafter
- ✓Perception (vision) encoder
- ✓Commercial use and redistribution permitted
- ✓No vendor support or SLA
Hosted inference via third-party providers
Usage-based (varies by provider)
- ✓Together AI, Fireworks AI and OpenRouter endpoints
- ✓Per-token metering set by each provider
- ✓No infrastructure to operate
- ✓Provider-dependent SLAs
The weights themselves are free under Apache 2.0 with no seat, token or revenue restrictions, so the real cost is hardware and operations: a 24GB consumer GPU is the practical floor for the K-Quant-17GB build and 32GB for K-Quant-Dynamic, while full BF16 precision needs 55GB or more. Meta publishes no list price and sells no support contract, so enterprises either absorb inference operations internally or pay per-token rates set independently by Together AI, Fireworks AI or OpenRouter. There is no enterprise tier to buy.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 5 on domains other than developer.meta.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Arena AI Evaluations
Paid model evaluations built on the Arena leaderboard's crowdsourced human votes, for enterprises choosing an LLM
Kimi K3
2.8T-parameter open-weight MoE model with a 1M-token context, priced at $3/$15 per million tokens
TypeSafe Jev
A 'System One' decision model that returns typed choices, scores and calibrated probabilities instead of text, for agent routing and classification
Xiaomi MiMo-V2.6
MIT-licensed, 1M-context multimodal MoE models that took the top open-weights score on Artificial Analysis at well under $1 per million tokens