Guide · Local AI

Muse Glimmer hardware requirements: what actually runs it

Published 2026-08-10 · fair values and eBay medians refresh daily · Methodology

Meta Superintelligence Labs released Muse Glimmer on August 10, 2026 under Apache 2.0. It is a distilled version of Muse Spark 1.2, aimed at agent workloads that stay running on your own machine instead of billing per token. Meta says it beats Gemma4-31B and Qwen3.6-27B on agentic and coding benchmarks including MCP Atlas and SWE-Bench Pro, and the weights are on Hugging Face with day-one support in Ollama, LM Studio, llama.cpp, MLX, and vLLM.

Meta did not size this one for a datacenter. They targeted a 24GB or 32GB memory envelope, which is exactly the used-GPU market this site tracks.

The memory math

Weights are only part of the bill. Muse Glimmer defaults to a 131,072-token context and can stretch to 262,144, and that KV cache lives in the same memory as the model. Meta's own guidance reserves headroom on top of the quantized weights for cache and the perception encoder, which is why a 17GB file wants a 24GB card rather than a 20GB one.

Paid links: we may earn a commission from eBay or Amazon purchases. Full disclosure.

GPU picks by tier

Fair values are our editorial midpoints. The eBay column is the live median Buy It Now across filtered samples, so it moves daily and sometimes sits above or below fair value.

The 3090 is the pick that keeps embarrassing newer cards. It carries the same 24GB as a 4090 for roughly $1,125 less, and a 30B model at 4-bit is bound by memory bandwidth rather than compute, so the gap in tokens per second is far smaller than the gap in price. Buy the 4090 if you also game or render. Buy the 3090 if this model is the job.

AMD owners are not locked out. The RX 7900 XTX has 24GB and a fair value near $900, which is the cheapest new-ish 24GB card on the board. ROCm support in llama.cpp works, though you will spend an evening on setup that CUDA users skip, and the speculative decoding path Meta tuned is a CUDA and Metal story right now.

Macs and the unified memory advantage

Apple Silicon shares one memory pool between CPU and GPU, so a 48GB Mac can hold a model that would need a workstation card on the PC side. Meta benchmarked M4 Max and M5 Max machines directly and reported a 1.5x decode speedup on M4 Max and 1.8x on M5 Max from DFlash speculative decoding. Those exact laptops are still new enough that the used market has not settled, so the value on our board sits in desktops.

The Mac mini M4 Pro with 48GB at $1,625 is the quiet winner here. It holds the 6-bit build, draws a fraction of the power of a 5090 rig, and costs less than half of one. The tradeoff is prompt processing: Apple's memory bandwidth is generous but its compute is not, so long prompts take noticeably longer to chew through before the first token appears. For a background agent that reads a repository once and then works, that matters less than it sounds. For interactive chat over huge contexts, it matters a lot.

Speculative decoding changes the card comparison

Meta ships a small drafter model alongside the main weights and calls the technique DFlash. It predicts several tokens ahead and lets the big model verify them in one pass, which produced 3.1x faster decoding on an RTX 5090, 1.8x on M5 Max, and 1.5x on M4 Max in Meta's numbers. The drafter needs its own slice of memory, which is another reason the 24GB floor is a floor rather than a comfortable fit.

That 3.1x figure is the strongest argument for the 5090 over a 3090 that has nothing to do with capacity. Two 3090s cost less than one 5090 and give you 48GB, but splitting a 30B dense model across two cards adds interconnect latency that eats into the win. If you want one box that does this well, the 5090 at $3,725 is the coherent answer. If you want the most VRAM per dollar and accept the tuning work, stack 3090s.

What to buy

Trying it out: a used RTX 3090 at $1,200, 4-bit, done. Running it as a daily agent on a machine you also use: a 48GB Mac mini M4 Pro at $1,625. Wanting the full quality tier and the speculative decoding speedup: an RTX 5090. Serving it to other people or running BF16: a Mac Studio M2 Ultra 128GB at $4,200, or workstation cards like the RTX A6000 at $3,750 in pairs.

One timing note before you buy. The Wall Street Journal reported that Meta plans to release open weights for Muse Spark 1.2 itself in the coming weeks, and that model is much larger than Glimmer. If your plan is to run the bigger one when it lands, buying for 24GB today means buying again later.

FAQ

How much VRAM does Muse Glimmer need?

24GB is the practical floor. The 4-bit quantization Meta ships is about 17GB of weights, and the rest goes to KV cache and the perception encoder — Meta describes the target as a 24GB or 32GB envelope. Full BF16 precision is 55.7GB and needs 58GB or more.

What is the cheapest GPU that runs Muse Glimmer?

A used RTX 3090. It has the same 24GB of VRAM as a 4090 at a fair value of $1,200 against $2,325, and for a memory-bound 30B model the older card gives up throughput rather than capability.

Can a Mac run Muse Glimmer?

Yes, and unified memory is the reason. Meta tuned the release for M4 Max and M5 Max machines. From our catalog the Mac mini M4 Pro with 48GB is the cheapest way to hold the 6-bit build at $1,625 fair value, and a Mac Studio M2 Ultra 128GB at $4,200 runs it unquantized.

Is Muse Glimmer free to use commercially?

The weights are Apache 2.0, which permits commercial use, modification, and redistribution without a separate license from Meta. That is a looser license than the Llama community license the older Meta models shipped under.

Do I need a 5090 or is a 3090 enough?

Enough depends on which quantization you want. Both cards run 4-bit. Only the 5090 and its 32GB holds 6-bit, and Meta measured a 3.1x decode speedup from DFlash speculative decoding on that card. You are paying about $2,525 more for the quality tier and the speed.

See all used GPU prices VRAM guide Mac vs GPU for local AI