Meta Muse Glimmer 30B Open Local Agent Model 4-bit Memory and DFlash [2026]

By Indie Kings | September 23, 2026

Updated September 23, 2026: Meta Muse Glimmer is a 30-billion-parameter open-weight model under Apache 2.0, built for local agents with tool use, multimodal input, long-horizon work, and failure recovery, per Meta's August 10 launch post. Meta says 4-bit weights fit below 20 GB and target 24 GB or 32 GB systems; NVIDIA separately reports more than 20,000 tokens per second per Blackwell Ultra GPU.

Open model launch graphic for Muse Glimmer

Image: Muse Glimmer launch graphic. Credit: NVIDIA Developer Blog.

What Meta released

Per Meta Superintelligence Labs, Muse Glimmer is a 30-billion-parameter model released with open weights under Apache 2.0. Meta positions it for always-on local workflows such as agents, function calling, coding, and LLM-as-a-judge evaluation, per the August 10 announcement.

Per Meta, the model is trained for long-horizon execution, precise tool calls, multimodal understanding, long-context memory, instruction following, and recovery after a failed tool call. It accepts interleaved text and images through a perception encoder, per the announcement.

Per Meta, the model supports controllable reasoning effort, works with OpenClaw and other scaffolds, and was trained on data from more than 100 languages. Meta says optimized integrations for llama.cpp, MLX, and ExecuTorch were coming in the following days, per the launch post.

Capability per Meta	Local-agent relevance
Tool use	Precise function calls across extended workflows
Long-horizon execution	Multi-step plans beyond one chat turn
Failure recovery	Diagnose failed calls and retry
Multimodal input	Screenshots, charts, and documents beside text
Controllable effort	Choose a quality and speed balance

Memory and quantization

Per Meta, full-precision Muse Glimmer would require more than 55 GB of memory. Meta uses quantization to compress weights to approximately 4-bit precision, reducing the language model below 20 GB, per the launch post.

Per Meta, the target 24 GB or 32 GB envelope leaves room for the KV cache, image perception encoder, and speculative decoding drafter. Weight size is not total runtime memory, so the envelope matters more than the headline parameter count, per the source's explanation.

Per Meta, the K-Quant-17GB model was measured with the quantized DFlash drafter on MacBook M4 Max, M5 Max, and RTX 5090 systems. Meta reports the setup as fluid for conversation and real-time agent interaction, per its own test description.

Memory item per Meta	Reported scale
Full-precision 30B model	More than 55 GB
4-bit language model	Below 20 GB
Target working envelope	24 GB or 32 GB
Extra runtime pieces	KV cache, perception encoder, drafter

DFlash speculative decoding

Per Meta, DFlash is a lightweight drafter model that proposes blocks of tokens while the main model verifies those proposals in parallel. Correct tokens are accepted and incorrect tokens are corrected, per the launch post.

Per Meta's chart, DFlash increases decode speed by 3.1 times on RTX 5090, 1.8 times on M5 Max, and 1.5 times on M4 Max in Meta's K-Quant-17GB test. These are Meta measurements, not an independent review, stated openly.

Meta test system	Reported DFlash gain	Source grade
RTX 5090	3.1 times decode speed	Meta measurement
M5 Max	1.8 times decode speed	Meta measurement
M4 Max	1.5 times decode speed	Meta measurement

Hardware and deployment paths

Per Meta, weights are available through Hugging Face, with paths for llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, Unsloth, vLLM, SGLang, and partner services. Meta says it is working with AMD, Arm, Dell, Intel, and NVIDIA on device optimization, per the announcement.

Per NVIDIA's developer post, Muse Glimmer is a dense 30B model with a reported context window above 120,000 tokens. NVIDIA lists RTX 5090, DGX Spark, DGX Station, and Jetson as deployment paths, per the article.

Per NVIDIA, Blackwell Ultra delivers more than 20,000 tokens per second per GPU at BF16 and NVFP4 precision in NVIDIA's setup. Treat that as a vendor benchmark tied to NVIDIA's runtime, model build, and hardware, not a universal guarantee.

Path	Source	Claim boundary
Hugging Face weights	Meta	Open-weight download
llama.cpp, MLX, ExecuTorch	Meta	Optimized integrations announced
vLLM, SGLang, NIM	Meta and NVIDIA	Serving and deployment paths
RTX 5090 and DGX	NVIDIA	Hardware-specific vendor results

What the claims do not prove

Meta says Muse Glimmer performs strongly against Gemma4-31B and Qwen3.6-27B in its size class, but the announcement does not provide an independent review or a single universal performance score in the text. Benchmark charts need workload, runtime, and quantization context, per the source.

Local inference can keep prompts and files on the device, but an agent scaffold can still grant tools and permissions. Model locality does not remove prompt-injection, credential, or tool-policy risk, so deployment needs the same care as any other agent.

No price, power draw, or total cost of ownership is claimed here. NVIDIA and Meta hardware figures describe their own test setups.

FAQ

What is Muse Glimmer?
Per Meta, a 30B open-weight model under Apache 2.0 for local agents, coding, tool use, and multimodal workflows.

How much memory does it need?
Per Meta, 4-bit weights take the language model below 20 GB, with a 24 GB or 32 GB target envelope for runtime pieces.

What is DFlash?
Per Meta, a lightweight drafter that proposes token blocks for main-model verification and faster decoding.

Can it run on one consumer GPU?
Meta says the model is sized for a single consumer GPU in quantized form. Actual fit depends on context, runtime, and tools.

What does NVIDIA report?
Per NVIDIA, more than 20K tokens per second per Blackwell Ultra GPU in its BF16 and NVFP4 setup.

Is it the same as Intel SuperClaw?
No. Muse Glimmer is Meta model weights. SuperClaw is Intel's separate agent application and routing system.

Bottom Line

Muse Glimmer makes local agents more practical by pairing open 30B weights with quantization and speculative decoding. Meta supplies the model baseline; NVIDIA supplies one hardware path. Test your own context, tools, and runtime before trusting a speed number.