Meta's new AI model runs at home now — one $1,299 AMD card almost keeps up with a $4,700 Nvidia one

Meta's Muse Glimmer 30B has been out for barely a day, and the hardware numbers are already in from both AMD and Nvidia. AMD says its $1,299 Radeon AI PRO R9700 hits 53 tokens per second on the model with its dFlash acceleration enabled, and that a Ryzen AI Max+ 395 laptop chip — no discrete GPU required — manages 24 tokens per second on its own. Nvidia's RTX 5090, a card that's been trading over $4,600 at retail this month, does 74.9 tokens per second stock and jumps to 233.4 with its own DFlash acceleration layer turned on. Same model, wildly different price-to-performance stories.

What Muse Glimmer actually needs

At full precision, Muse Glimmer wants 64GB of VRAM — out of reach for almost anyone outside a workstation. Quantized down with AMD and Meta's K-Quant-Dynamic format, that drops to roughly 32GB with only a 0.2% accuracy loss. Push it further to the K-Quant-17GB variant and it fits in 24GB, at a still-modest 1.0% accuracy loss. It's a 30-billion-parameter model with a 131,000-plus-token context window, Apache 2.0 licensed for commercial use, and it takes both text and images as input — background we covered when Meta open-sourced the model as a direct swing at OpenAI and Anthropic's closed offerings.

The AMD numbers

The Radeon AI PRO R9700 is a 32GB GDDR6 workstation card built on the same Navi 48 RDNA4 die as the gaming RX 9070 XT, just with double the memory and a 300W power target. At $1,299 MSRP, it's the cheapest card AMD sells with enough VRAM to comfortably run Muse Glimmer's 32GB quant with headroom to spare. On the laptop side, the Ryzen AI Max+ 395 — an APU, not a discrete GPU — pulls it off using unified system memory instead, through LM Studio or plain llama.cpp. AMD's own guidance is straightforward: 32GB or more of VRAM, or Variable Graphics Memory on the APU side, and you're in.

Where the $1,299 card actually wins

Do the rough math and the R9700 is the better deal right up until you turn Nvidia's acceleration on. At stock speeds, you're paying roughly $17 per token/second on the AMD card versus around $61 per token/second on an RTX 5090 bought at this month's inflated retail price. Flip DFlash on for the 5090, though, and that gap collapses — Nvidia pulls ahead in raw throughput while costing under $20 per token/second at the new, higher rate. This is my read, not an official benchmark: if you're doing anything latency-sensitive — coding assistance, multi-turn agents — the 5090's ceiling matters more than its price. If you're running batch jobs overnight or don't need instant responses, the R9700 is doing 70% of the accelerated Nvidia card's unaccelerated speed for well under a third of the cost.

Nvidia's answer: DFlash

The single biggest number in either company's numbers is that 3.1x jump on Nvidia's side — 74.9 tokens per second stock to 233.4 with DFlash enabled. AMD's dFlash gets credit for the R9700's 53 tokens per second figure too, though AMD hasn't published an unaccelerated baseline to compare against directly. For context on the low end, Apple's M5 Max manages 26.6 to 50.2 tokens per second depending on configuration, and the M4 Max runs 23.7 to 37.8 — putting even a top-end MacBook in roughly the same range as AMD's laptop APU, not its discrete workstation card.

Expect these numbers to keep moving. Both companies are actively tuning their acceleration stacks, and it wouldn't be surprising to see AMD publish an unaccelerated dFlash-off baseline soon to make the comparison cleaner. For now, the takeaway holds: a $1,299 card gets you into this model comfortably, and you don't need Nvidia's most expensive GPU to make good use of what Meta just shipped.