Present
Always-on workflows without depending on cloud availability.
Open weights · Apache 2.0
An agentic model
that stays with you.
Muse Glimmer brings long-horizon reasoning, precise tool use, and multimodal understanding to the computer in front of you.
30Bparameters
<20 GBquantized language model
100+training languages
Muse Glimmer / Meta Superintelligence Labs / August 10, 2026
Why local agents
The cloud made AI powerful.
Your device makes it personal.
An agent that manages schedules, drafts messages, organizes files, and learns how you work needs access to personal context. Running locally can keep the model available with or without a network connection, close to the tools and information it acts on.
Responsive
Designed around practical latency on consumer hardware.
Open
Weights released under the permissive Apache 2.0 license.
Inside an agentic workflow
Observe. Decide. Act.
Check the world again.
Agentic intelligence is not one answer. It is a continuing loop between what the model sees, the plan it forms, the tool it calls, and the result it verifies.
Conceptual visualization — not a measured result or literal model architecture
Your device local context
01ObserveCalendar shows two meetings at 10:00.
02ReasonCompare priority, attendees, and free time.
03ActCall the calendar tool with a precise update.
04VerifyRead the tool result; recover if it conflicts.
Muse
Glimmer
Glimmer
Step 1 of 4 The perception encoder can interpret interleaved text and images.
Built for agents
Capabilities that work
in concert.
An effective agent has to hold a plan, call the right tools, interpret what comes back, and recover when the world does not match expectations.
End-to-end task completion
Trained and evaluated on full tasks across research, code, and multi-turn workflows.
Reliable tool use
Invokes functions with precise schemas throughout extended workflows.
Failure recovery
Diagnoses unexpected tool results and retries rather than simply halting.
Multimodal reasoning
A dedicated perception encoder accepts interleaved text and images, including screenshots, charts, and documents.
Controllable effort
Supports different reasoning strengths to balance quality and speed.
Also trained for multi-step reasoning, scaffold compatibility across OpenClaw and other orchestration patterns, and multilingual use across more than 100 languages.
Training recipe
Capability,
compressed.
Muse Glimmer balances agentic performance against the memory and compute constraints of local hardware.
Learn from a larger teacher.
Logit distillation from Muse Spark outputs, using a similar data mix as the teacher.
Extend the horizon.
Longer-context, more agent-heavy data with richer reasoning traces, alongside organic data.
Practice the whole loop.
Supervised fine-tuning, on-policy distillation, and reinforcement learning across general, reasoning, coding, and agentic domains.
Optimized for local deployment
Make 30B fit.
Then make it feel fast.
01 / QuantizationSource-reported
From over 55 GB
to under 20 GB.
Approximately 4-bit weights leave headroom for KV cache, the perception encoder, and the speculative decoding drafter within a 24 GB or 32 GB envelope.
Full precision>55 GB
Quantized LM<20 GB
Meta reports minimal to no degradation on agentic tasks from this compression.
02 / Speculative decodingInteractive concept
Draft in blocks.
Verify in parallel.
A lightweight DFlash-based drafter proposes token blocks. The main model accepts correct tokens and corrects wrong ones while preserving identical output quality.
Callthecalendartoolwiththenewtime.
accepted draft corrected
Measured decode speedup
More tokens through the loop.
K-Quant-17GB with the quantized DFlash drafter, measured by Meta on local hardware.
Read the source methodologyBroad evaluation
One agent, many demands.
Meta evaluated Muse Glimmer across agentic task completion, coding, multimodal understanding, safety, reasoning, instruction following, and LLM-as-a-judge use. Reported comparisons include similarly sized Gemma4-31B and Qwen3.6-27B.
- DeepSearch QA
- MCP-Atlas
- τ-Bench
- SWE-Bench
Open weights, available now
Put an agent
where the work is.
Download Muse Glimmer’s weights, review the developer documentation, and choose the runtime that fits your workflow.
Run locally
Ollama · LM Studio · Unsloth
Integrations announced for the coming daysEdge frameworks
llama.cpp · ExecuTorch · MLX
Optimized integrations announcedServe at scale
vLLM · SGLang
Source-listed serving optionsOptimize
AMD · Arm · Dell · Intel · NVIDIA
Hardware partners named by Meta