What hardware do you need to run Muse Glimmer 30B locally?

Verified Aug. 25, 2026: Meta introduced Muse Glimmer 30B on Aug. 10, 2026, but the phrase that matters for buyers is not the launch-friendly “under 20 GB” compressed-weight line. The model card points to a roughly 24 GB memory target for the official K-Quant-17GB release, 32 GB for the higher-fidelity K-Quant-Dynamic release, and 64 GB for BF16; it also says audio input and output are not supported [1][2]. For a local smart-home assistant, that means the real requirement is the model envelope plus Home Assistant integration plus separate speech-to-text and text-to-speech, not just the language model file.

The memory table that should decide the purchase

The easiest way to buy the wrong machine is to treat the main GGUF file as the whole assistant. In the current local release, the 30B language model is only one artifact. The vision projector and the DFlash drafter are separate pieces, and the system still needs runtime memory for context, Home Assistant calls, and whatever speech stack you bolt on.

That table is why “fits under 20 GB” is a bad buying rule. The main file, projector, and drafter already total about 19.79 GB before context cache, runtime overhead, Home Assistant integration, and speech components are considered [3][4]. A 24 GB card can be a real floor for the official compressed release, but it is not a roomy floor.

It also explains why the 32 GB tier is not just vanity. The K-Quant-Dynamic variant is published with a lower average degradation figure than K-Quant-17GB, and the extra memory gives the rest of the assistant somewhere to live [2]. In a smart-home setup, the machine is not only writing a reply. It may be holding recent household context, parsing a tool call, waiting on Home Assistant, recovering from a failed device call, and preparing spoken output.

What actually qualifies: 24 GB, 32 GB, 64 GB

For the official compressed Muse Glimmer 30B setup, I would treat 24 GB of usable GPU memory or unified memory as the entry point. That does not mean every 24 GB machine will feel good with every context length and every voice pipeline. It means the machine is at least in the right conversation.

If you already own a 24 GB card

An RTX 3090 or RTX 4090 is the most defensible “try it before buying anything else” tier. Use the official compressed release, keep context conservative, and do not assume the speech stack can take whatever memory is left. A used RTX 3090 can be the value play if the system has the power supply, cooling, and case space for it. An RTX 4090 is faster and easier to live with thermally in many builds, but it does not change the 24 GB ceiling.

If you are buying new for Muse Glimmer

The 32 GB class is the cleaner target if the machine is being bought for this model rather than repurposed. Meta’s own vendor-reported DFlash numbers include an RTX 5090 run, and AMD has published preliminary results for Radeon AI PRO R9700 32 GB systems, including a 53 tok/s figure in its own testing [1][5]. Those speed numbers are useful signals, not guarantees for your Home Assistant box. Driver stack, backend, context length, request format, and whether speech is running on the same machine all matter.

The Radeon AI PRO R9700 is interesting because 32 GB is exactly the memory class this model wants for a less cramped assistant. The caution is not about the memory size; it is about ecosystem friction. If your local AI stack is already CUDA-shaped, a 32 GB NVIDIA path may be less work. If your workload and OS fit AMD’s current tooling, the memory math is attractive.

Mac unified memory is viable, but it is still shared memory

The model card lists M4 Max and M5 Max Macs among supported local-device targets, and local runners such as Ollama and MLX-style paths make Apple silicon a practical option for people who want one quiet machine rather than a gaming tower [2]. The important correction is that 32 GB unified memory is not the same operating experience as 32 GB of empty VRAM. macOS, the runner, the browser tab you forgot about, Home Assistant add-ons, and speech components all share the same pool.

A 32 GB Mac can be a comfortable compressed-model tier if you keep the rest of the stack tidy. A 64 GB Mac lines up better with full-precision ambitions, longer context, and running more of the assistant on the same box. If the Mac is also your daily workstation, buy more memory than the model table suggests.

Jetson and appliance-style boxes need a different expectation

NVIDIA-affiliated Jetson AI Lab materials report Muse Glimmer 30B throughput in the 25–36 tok/s range on Jetson-class systems [6]. That is promising for edge deployments, especially if the goal is a contained local appliance. It should not be read as the same buying decision as a desktop GPU. Embedded platforms have their own memory sharing, quantization, thermal, and software constraints.

Quantization: loading once is not the same as living with it

Community quant ladders make the model look much more flexible than the official release table. Unsloth lists options from roughly 2-bit quants in the 12–14 GB range up to BF16 around 58 GB [3]. Those numbers are useful if you like squeezing models into marginal hardware, but they are not a replacement for an assistant requirement.

For a dashboard toy, a very low-bit quant might be fine. For a household assistant, the penalty is paid in the annoying places: tool-call reliability, instruction following, recovery after a failed device action, and the model’s ability to keep room names, entity names, and user preferences straight. If a model says it is turning off the “kitchen pendant” but calls the wrong entity, nobody in the house cares that the quant loaded successfully.

That is why the official K-Quant-17GB release is the practical floor, not the heroic floor. The K-Quant-Dynamic tier earns attention because its published degradation figure is lower and its memory target gives more operating space [2]. BF16 is the cleanest representation, but it pushes the build into 64 GB territory [2].

Context: 131K exists, but do not plan a 24 GB assistant around it

Muse Glimmer’s maximum context is published as 131K tokens [2]. That is not the context target I would use for a 24 GB smart-home build. Long context consumes memory through the KV cache, and the assistant still needs room for the loaded model, drafter, runtime overhead, tool responses, and speech components.

Sebastian Raschka’s architecture notes estimate Muse Glimmer’s grouped-query-attention KV cache at about 52 KiB per token, compared with 64 KiB for Qwen3.6-27B and 840 KiB for Gemma 4 31B [7]. On paper, that makes Muse Glimmer much more long-context-friendly than some alternatives. In a 24 GB box, though, a planning target around 32K tokens is the saner place to start. At roughly 52 KiB per token, 32K context is already about 1.6 GiB of KV cache before allocator behavior and other runtime costs are considered; 131K is about 6.5 GiB.

A smart-home assistant rarely needs to carry a whole novel of context to turn off a light. It needs the current request, recent dialogue, the relevant device schema, a few rules about the household, and enough history not to repeat a mistake. Spend memory on reliability before spending it on a maximum-context trophy.

Will it behave like a smart-home assistant?

Muse Glimmer is more interesting than a plain chat model for this job because Meta trained and presented it as an agentic local model. Meta reports MCP Atlas 75.5 and DeepSearch QA 74.6 results, and describes failure-recovery training meant to help the model recover when an action or tool call goes wrong [1]. Those are vendor-published benchmark and training claims, not proof that your kitchen lights will behave. They are still relevant because smart-home control is mostly tool use with consequences.

In Home Assistant terms, the useful path is not magic. You run Muse Glimmer behind a local model server, expose it through something Home Assistant can call, and make sure the model receives a controlled list of entities and services instead of the whole house dumped into the request context. Ollama already lists a muse-glimmer model entry, and Home Assistant-focused projects such as home-llm support local LLM control through local or OpenAI-compatible endpoints [8][9].

If you are comparing local control with cloud assistants, the practical question is similar to the one in our Grok smart-home control check: what can actually call Home Assistant, what stays local, and what happens when the assistant misunderstands a command. Muse Glimmer changes the privacy and latency equation because it can run on your device, but it does not remove the need for guardrails around service calls.

Voice is a separate system, not a hidden feature

The model card’s audio line deserves more attention than it will get on most launch charts: audio input and output are not supported [2]. A voice smart-home assistant therefore needs a speech-to-text component before Muse Glimmer and a text-to-speech component after it. Faster-whisper and Piper are common local choices, but they are separate parts of the bill of materials.

This matters for hardware because speech is not free. If STT runs on the same GPU, it competes for memory with the model and context. If it runs on CPU, it may add latency. If TTS runs locally, it needs its own model, process, and audio plumbing. None of those pieces changes the Muse Glimmer weight size, but all of them change whether the assistant feels responsive in a kitchen or hallway.

No source I would trust today documents a complete Muse Glimmer 30B plus Home Assistant plus local STT/TTS pipeline with end-to-end latency numbers across common hardware. So the honest requirement cannot be “buy X and voice will be fast.” The honest requirement is: leave memory and compute for speech, or put speech on a different device.

Speed numbers: encouraging, but label the provenance

DFlash is the part of the release that could make Muse Glimmer feel less like a research model and more like a usable assistant. Meta reports speculative-decoding gains from 74.9 to 233.4 tok/s on RTX 5090, a 3.1x increase, and from 23.7 to 37.8 tok/s on M4 Max, a 1.5x increase [1]. That is the right kind of speedup for a voice interface, where a slow first response makes the whole setup feel broken.

The restraint is provenance. Meta’s DFlash results are Meta-published. AMD’s 53 tok/s Radeon AI PRO R9700 number is vendor-preliminary [5]. Jetson’s 25–36 tok/s range comes from an NVIDIA-affiliated lab [6]. Kingy and Wavect both present useful hardware and benchmark guides, but the crawled materials also warn that many launch-window numbers had not been independently reproduced as of August 2026 [4][10].

That does not make the numbers worthless. It just means they should not be converted into a shopping guarantee. A model that is fast in a batch-one greedy-decoding test can still feel sluggish if your Home Assistant request is bloated, your entity list is unfiltered, your context window is too large, or your STT process is fighting the LLM for memory.

Buy, keep, or skip

Keep and test a 24 GB RTX 3090 or RTX 4090 if you already own one. Run the official compressed path, start around a practical context target instead of the 131K maximum, and be conservative with speech. This is the floor I would trust for tinkering toward a real local Home Assistant brain.

Buy 32 GB if Muse Glimmer is the reason for the purchase. That tier better matches the K-Quant-Dynamic target, gives breathing room for context and Home Assistant tooling, and makes the voice pipeline less likely to turn into a memory juggling act. RTX 5090-class and Radeon AI PRO R9700-class systems belong here, with the normal caveat that software support may matter as much as raw memory.

Choose 64 GB or more if you want BF16, long-context experiments, heavier multimodal use, or multiple local AI services running beside Home Assistant. This is also the tier I would look at for a Mac that doubles as a daily machine, because unified memory is shared with everything else the system is doing.

Skip buying a 16 GB GPU or a marginal mini PC specifically for Muse Glimmer 30B as a local smart-home assistant. You may get a low-bit quant to load, and that can be fun. It is not the same as running the official release variant at a useful context length while leaving room for Home Assistant and local voice.

The responsible requirement for Muse Glimmer 30B is the whole assistant envelope: official model variant, projector and drafter if used, context cache, Home Assistant integration, and separate STT/TTS. The compressed-weight number is only the first box on the desk.

References

1. Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device, Meta AI, Aug. 10, 2026
2. meta-models/Muse-Glimmer-30B Model Card, Hugging Face
3. Muse Glimmer - How to Run Locally, Unsloth
4. Muse Glimmer 30B: Benchmarks, Hardware & How to Run, Kingy AI
5. Run Meta Muse Glimmer 30B on AMD Ryzen AI Max Agentic PCs and Radeon GPUs, AMD
6. Muse Glimmer 30B | Jetson AI Lab, Jetson AI Lab
7. Muse Glimmer 30B Architecture Notes, Sebastian Raschka
8. muse-glimmer | Ollama Library, Ollama
9. acon96/home-llm, GitHub
10. Muse Glimmer 30B Hardware and Benchmark Guide, Wavect