Computer Forum
Meta Muse Glimmer Release (Image © Meta)
Memory Optimization Through Distillation
The standard operation of a model with 30 billion parameters at fp16 precision requires approximately 60 GB of memory for the weights alone. To circumvent this hardware limitation, Meta employed distillation from a larger model called “Muse Spark.” This process reduced memory requirements to 20 GB. Combined with the 2 GB to 4 GB required for the KV cache, the total memory requirement falls within the limits of consumer-grade graphics cards with 24 GB or 32 GB of VRAM. Meta reports that this compression method does not result in any noticeable loss of performance.
Token Generation Speed and Drafting
The model uses a companion system known as the “DFlash Drafter model” to reduce response latency. This smaller model predicts text segments, which “Muse Glimmer” then verifies in a single pass. Since verifying a response is faster than generating one from scratch, this architecture increases output speed.
On a GeForce RTX 5090 (Test) graphics card, responses are 3.1 times faster. Improvements were also seen on Apple hardware; the M5 Max and M4 Max recorded speed increases of 1.8x and 1.5x, respectively.
Global Market Context and Usage Trends
The timing of this release coincides with a period of significant growth for Chinese AI models. Data from OpenRouter shows that Chinese architectures have led the global token count for fifteen consecutive weeks.
In the week beginning August 3, total global usage of AI models reached 69 trillion tokens, representing a weekly increase of 21.48 percent. Of this total, 34.25 trillion tokens were attributed to Chinese models, while U.S. models contributed 9.17 trillion. In particular, DeepSeek V4 Flash recorded a week-over-week growth rate of 570 percent.
