TL;DR: AMD shipped day-zero support for Meta’s Muse Glimmer 30B, and it published real numbers: up to 24 tok/s on a Ryzen AI Max+ 395 mini PC and up to 53 tok/s on the $1,299 Radeon AI PRO R9700 — both measured with dFlash speculative decoding enabled, in llama.cpp on the Vulkan backend. The quiet headline is that dFlash is not a CUDA exclusive. The loud caveat is that an RTX 5090 does 233 tok/s on the same model.

	Ryzen AI Max+ 395 (Strix Halo)	Radeon AI PRO R9700 32GB	Used RTX 3090 24GB
Best for	Capacity: full dFlash stack + big context in 64GB VGM	Cheapest new 32GB card that fits model + drafter + cache	Fastest base-model decode per dollar
Price / Cost	~$1,999+ as a 128GB mini PC, DRAM-crisis volatile	$1,299 MSRP, $1,522–$2,139 street	$1,248 avg used (Aug 2026)
Measured speed	24 tok/s with dFlash	53 tok/s with dFlash	~65–70 tok/s base (bandwidth estimate)
The catch	256 GB/s bandwidth is the floor of viability	Prefill still 2.6–3.4× slower than NVIDIA	No room for the 5.11GB drafter

Honest take: if you already own the 24GB NVIDIA card, nothing here changes your setup — run the base model and keep your ~70 tok/s. But if you’re buying hardware specifically to run the full Glimmer stack — model, dFlash drafter, vision encoder, and real context, all resident — the R9700 at $1,299 is now the cheapest new card that does it, at less than a third of an RTX 5090’s price and a bit under a quarter of its dFlash speed.

Meta’s Muse Glimmer 30B launched on August 10 with NVIDIA and Apple benchmarks front and center — we covered the general GPU-tier math in our hardware guide. One day later, AMD published its own day-zero enablement post covering Ryzen AI Max+ systems and the Radeon AI PRO R9700, with measured tokens-per-second and a named software stack. This article is the AMD-specific follow-up: what AMD actually demonstrated, what the numbers mean against the NVIDIA and Apple figures, and which AMD buyers this model is genuinely for. Check your own card’s fit first with the VRAM calculator.

What AMD shipped on day zero

Three concrete things, per AMD’s blog and AMD’s launch post on X:

1. llama.cpp Vulkan support, with dFlash working. AMD’s benchmarks were run on Windows with the Vulkan backend and dFlash speculative decoding enabled, using --spec-draft-n-max=4 on a preliminary Muse Glimmer build. That answers the question we flagged as critical in the general guide: Meta’s block-diffusion drafter is not locked to CUDA or ExecuTorch. It runs through the same open llama.cpp path AMD cards already use.
2. LM Studio integration. LM Studio lists Muse Glimmer and runs it out of the box on supported AMD systems with more than 32GB of memory — which in practice means Ryzen AI Max+ machines and the R9700, exactly the two products AMD named.
3. Lemonade catalog integration. Lemonade Server added Muse Glimmer to its catalog in version 11.6, pulling the Unsloth UD-Q4_K_XL GGUF — a 15.9GB weights file, roughly 20GB downloaded with the vision projector — and serving it over OpenAI-, Ollama-, and Anthropic-compatible APIs. If you want Glimmer behind a local API endpoint on AMD without touching llama.cpp flags, this is the shortest path. We covered Lemonade’s broader stack in our AMD Lemonade guide.

Ollama works too — v0.32.8 added AMD support after the Apple-only v0.32.7 launch window — but AMD’s own demonstrations ran llama.cpp and LM Studio, and that choice is informative: Vulkan is the path AMD validated.

The numbers, in context

AMD’s two measurements, alongside the Meta-published NVIDIA and Apple figures from the launch:

Hardware	Memory / bandwidth	Muse Glimmer 30B speed	Source, conditions
RTX 5090 32GB	1,792 GB/s	74.9 → 233.4 tok/s with dFlash	Meta, llama.cpp, batch 1
RTX 4090 24GB	1,008 GB/s	75 tok/s base, no drafter	Independent single-card bench
Radeon AI PRO R9700 32GB	640 GB/s	up to 53 tok/s with dFlash	AMD, llama.cpp Vulkan, Windows
Apple M5 Max	546 GB/s class	26.6 → 50.2 tok/s with dFlash	Meta, ExecuTorch
Apple M4 Max	546 GB/s	23.7 → 37.8 tok/s with dFlash	Meta, ExecuTorch
Ryzen AI Max+ 395	256 GB/s	up to 24 tok/s with dFlash	AMD, llama.cpp Vulkan, Windows, VGM 64GB
Used RTX 3090 24GB	936 GB/s	~65–70 tok/s base (our bandwidth-ratio estimate)	No clean public bench yet

Two caveats before reading too much into a single table. These are vendor numbers measured on different stacks — Meta used ExecuTorch for Apple and llama.cpp for the 5090; AMD used a preliminary model build in llama.cpp Vulkan — so treat cross-vendor gaps as directional, not decimal-precise. And AMD says “up to,” which in our experience of RDNA4 llama.cpp benchmarking means a tuned driver and a short prompt.

With that said, the pattern is exactly what the bandwidth column predicts, because Glimmer is a dense 30B and dense decode follows memory bandwidth. The R9700’s 640 GB/s lands it between the M4/M5 Max and the 24GB GeForce cards. Strix Halo’s 256 GB/s puts it at the bottom of the viable range — 24 tok/s is comfortably above the 7–10 tok/s reading-speed floor, but it’s the floor of comfortable, not the middle.

What dFlash is worth on AMD. Here’s the derivation the launch posts skip. The R9700 decodes dense ~27B-class Q4 models at 29–33 tok/s without any drafter — we published those measurements in our R9700 review, and the bandwidth math (a ~17GB weight sweep through 640 GB/s ≈ 37 tok/s ceiling) agrees. So 53 tok/s with dFlash implies a ~1.6–1.8× effective speedup on RDNA4 — right in line with Apple’s 1.5–1.8×, and well short of the 5090’s 3.1×. That’s not a knock on AMD’s port: speculative decoding shifts work from bandwidth-bound token-by-token decoding to compute-bound block verification, and compute throughput is precisely where Blackwell’s lead over RDNA4 is widest — the same reason the R9700’s prefill trails NVIDIA by 2.6–3.4× in our review. dFlash helps AMD; it helps NVIDIA more.

The NPU does nothing here. A Ryzen AI Max+ 395 has three compute engines — 16 Zen 5 cores, a 40-CU RDNA 3.5 iGPU, and a 50 TOPS XDNA 2 NPU — and AMD’s own Glimmer demo uses exactly one of them. Vulkan is an iGPU path; the NPU sits idle, consistent with everything we found in our NPU vs GPU analysis: for bandwidth-bound LLM decode, TOPS are the wrong currency, and no mainstream llama.cpp path drives the XDNA NPU for a 30B dense model. If you were hoping “Ryzen AI” meant the NPU had finally found its LLM job, this launch is another data point that it hasn’t.

Where AMD actually wins: the memory math

Speed is NVIDIA’s column. Capacity is AMD’s, and Glimmer’s stack makes the capacity story unusually concrete, because running everything takes real memory:

• Q4-class weights: 15.9–17.3GB depending on quant
• dFlash drafter: +5.11GB
• Vision/perception encoder: loaded separately, on top
• KV cache: small for this model (a 130K context fit in 19.3GB total on a 4090), but not zero

On a 24GB GeForce card, that stack doesn’t fit — that’s the dFlash catch we documented at launch, and it’s why the independent 4090 number in the table is a base number. The two AMD products AMD chose to demo are, not coincidentally, the two configurations where the whole stack fits:

Radeon AI PRO R9700 — the cheapest new card that runs all of it. 32GB swallows weights + drafter + encoder + cache with room left over. The R9700 listed at $1,299 MSRP ($1,522–$2,139 street in August); the only other new 32GB consumer-adjacent card is the RTX 5090 at $4,399-plus. Our 32GB tier guide already called the R9700 the tier’s value pick — Glimmer’s full stack is now the clearest single workload that justifies it. You give up half the dFlash speed; you keep $3,000.

Ryzen AI Max+ 395 — when you want the model to be one of several things running. AMD’s test system was configured with 128GB of system memory and Variable Graphics Memory set to 64GB. At 64GB of GPU-addressable memory, Glimmer’s whole stack occupies about a third of the budget — leaving room for the K-Quant-Dynamic tier (Meta’s higher-quality 32GB-target quant), a second resident model, or a genuinely long-lived agent session, scenarios we mapped in the Strix Halo deep dive. A GMKtec EVO-X2 with 128GB listed at $1,999 at Micro Center in early August, but DRAM-crisis pricing on 128GB machines is moving week to week — check before you order, and don’t pay 24GB-GPU money for a machine you only want for a model that fits a 24GB GPU.

What about the Radeon cards people actually own? AMD’s post names only the R9700 on the discrete side. An RX 7900 XTX’s 24GB and 960 GB/s should run the base model at roughly used-3090 speeds — call it ~65 tok/s as a bandwidth-ratio estimate, with the same no-room-for-the-drafter constraint, but no one has published a clean XTX measurement yet and AMD conspicuously didn’t. The 16GB cards (RX 9070 XT, RX 7800 XT) are out for the Q4 tier: 15.9–17.3GB of weights doesn’t fit 16GB, and a dense 30B split across PCIe decodes in single digits. On 16GB, the 2-bit quants technically load — Unsloth’s own repo shows a 2-bit Glimmer making 100+ tool calls in 14GB of VRAM — but for daily use the 16GB tier’s native picks remain the better trade.

The problem you’ll actually hit: VGM on Strix Halo

The most common failure mode on Ryzen AI Max machines won’t look like a failure — it’ll look like a dense 30B running at 4 tok/s. Windows doesn’t hand the iGPU 64GB by default; out of the box, a 128GB Strix Halo machine typically reserves a much smaller dedicated slice, and llama.cpp offloads what fits into that slice and quietly runs the rest on CPU:

$ llama-server -m Muse-Glimmer-30B-UD-Q4_K_XL.gguf -ngl 99
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon(TM) 8060S Graphics ... free memory 30720 MiB
...
load_tensors: offloaded 34/49 layers to GPU     ← the tell: not 49/49

The fix is the setting AMD’s own test config names: open Adrenalin → Performance → Tuning → Variable Graphics Memory, set it to 64GB (on a 128GB machine), reboot, and reload — the log should then show every layer offloaded, and full-precision-feeling speed comes back. Same class of silent-CPU-spill bug we keep finding on NVIDIA setups; AMD’s version just hides in a driver panel instead of an environment variable.

On the software-path question — Vulkan vs ROCm — the practical answer in August 2026 is: use what AMD used. Vulkan is the demonstrated, Windows-friendly path and our RDNA4 Vulkan vs ROCm benchmarks show it’s no longer the slow option; ROCm 7.2 on Linux works for the R9700 if you’re already living there. There is no CUDA-style single blessed path, and for this model there doesn’t need to be one.

Who should buy AMD for Muse Glimmer

• You want the full agentic stack — dFlash, vision, long context — on a new card without RTX 5090 money: the R9700 is now the default answer at $1,299 MSRP. Budget for the prefill gap if your agent slings 50K-token prompts.
• You want Glimmer as one resident service among several on an always-on box: Ryzen AI Max+ 395 with 128GB, VGM at 64GB. Accept 24 tok/s; buy it for the memory ceiling, not the speed — the same verdict as our EVO-X2 review.
• You own a 24GB card of either color: change nothing. Base model, ~65–75 tok/s, full context, drafter skipped — still the best speed-per-dollar way to run this model, and the used RTX 3090 at $1,248 average remains the cheapest ticket into it.
• You want dFlash’s 233 tok/s ceiling or you’re prefill-bound: that’s NVIDIA’s column, and it costs $4,399+. Rent it first — an hour on a 32GB+ GPU via RunPod will tell you whether your agent loop actually feels different at 233 tok/s before you commit four grand, per the math in rent vs buy.

If you’re wiring Glimmer into a coding agent on any of this hardware, aicoderscope.com covers the tool side; for self-hosted serving stacks around it, see aifoss.dev.

FAQ

Does dFlash require CUDA? No. AMD’s published numbers were measured with dFlash enabled in llama.cpp on the Vulkan backend, on both Strix Halo and the R9700. The speedup is smaller on RDNA4 (~1.6–1.8× effective vs 3.1× on an RTX 5090) because block verification is compute-bound, but it works.

Does the Ryzen AI Max NPU accelerate Muse Glimmer? No. AMD’s demonstrated path is llama.cpp Vulkan, which runs on the iGPU. The XDNA 2 NPU is not used for this model on any mainstream inference stack as of August 2026.

Can an RX 9070 XT (16GB) run it? Not at usable quality. The Q4-class weights alone are 15.9–17.3GB. Two-bit quants fit in ~14GB but are the quality floor for a dense 30B; see the 16GB tier guide for what 16GB runs well instead.

Is the R9700 faster than a used RTX 3090 for this model? For base-model decode, no — the 3090’s 936 GB/s should put it around 65–70 tok/s (estimate) vs the R9700’s ~30 tok/s class without the drafter. The R9700’s case is the 32GB: it runs the model with the dFlash drafter and vision encoder resident at 53 tok/s, which no 24GB card can do at all.

What’s the fastest way to serve it on AMD with an API? Lemonade Server 11.6+ — it pulls the Unsloth UD-Q4_K_XL GGUF from its catalog and exposes OpenAI-, Ollama-, and Anthropic-compatible endpoints locally. LM Studio (0.4.x, AMD systems with >32GB memory) is the GUI alternative.

Recommended Gear

• Radeon AI PRO R9700 32GB — cheapest new card that runs the full Glimmer stack
• GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB) — the capacity play; verify current pricing first
• Used RTX 3090 24GB — best base-model speed per dollar
• RX 7900 XTX 24GB — the unofficial AMD 24GB option, base model only

Sources

• Run Meta Muse Glimmer 30B on AMD Ryzen AI Max Agentic PCs and Radeon GPUs — AMD
• AMD Expands Local AI Capabilities With Muse Glimmer 30B Support — TipRanks
• Meta’s Muse Glimmer Hits 24 Tokens/Sec on AMD’s New AI Chip — Techgenyz
• Introducing Muse Glimmer — Meta AI Research
• AMD day-zero support announcement — AMD on X
• Run Meta’s Muse Glimmer 30B locally with Lemonade — Lemonade Server
• Lemonade 11.6 Integrates Muse-Glimmer 30B — Phoronix
• unsloth/Muse-Glimmer-30B-GGUF — Hugging Face
• meta/muse-glimmer — LM Studio model catalog
• AMD Radeon AI PRO R9700 specs and pricing — GPU Poet
• AMD unveils its $1,299 Radeon AI PRO R9700 32GB workstation GPU — OC3D
• GMKtec EVO-X2 AI Mini PC 128GB — Micro Center

Last updated August 15, 2026. Prices and specs change; verify current rates before purchasing. AMD’s performance figures are vendor-published preliminary results (“up to,” Windows, llama.cpp Vulkan, dFlash enabled); cross-vendor comparisons in this article mix test stacks and should be read as directional.