Meta has released Muse Glimmer, a 30 billion parameter open weight AI model designed to run locally on high end consumer hardware while supporting coding and agent based workloads.
The model is notable for fitting within roughly 20GB of memory for its weights, allowing it to run on graphics cards with 24GB or 32GB of VRAM when additional memory for the KV cache is included.
Meta is also using a smaller companion model called DFlash to accelerate token generation. According to the supplied figures, this approach can improve response speed by 3.1 times on an NVIDIA GeForce RTX 5090.
Muse Glimmer key details
| Feature | Detail |
|---|---|
| Model size | 30 billion parameters |
| Model type | Open weight |
| Target workloads | Local agents, coding and multistep tasks |
| Approximate model weight memory | 20GB |
| Additional KV cache requirement | Around 2GB to 4GB |
| Suitable GPU memory | 24GB or 32GB |
| Companion model | DFlash |
| RTX 5090 response speed improvement | 3.1x |
| Apple M5 Max improvement | 1.8x |
| Apple M4 Max improvement | 1.5x |
Muse Glimmer is positioned against other open models in the same general class, including Google’s Gemma 4 and Alibaba’s Qwen 3.6.
Model compression helps Muse Glimmer fit on one GPU
A 30 billion parameter model running at FP16 precision would normally need about 60GB of memory for its weights alone.
Meta has reduced that requirement substantially.
The company used quantization along with a distillation process in which a larger model, Muse Spark, trains Muse Glimmer. The goal is to retain much of the larger model’s capability while reducing memory requirements.
The resulting model weights reportedly use around 20GB.
With another 2GB to 4GB needed for the KV cache, the total requirement becomes manageable for GPUs with 24GB or 32GB of VRAM.
That means cards in the high end consumer market can potentially run the model without needing multiple GPUs or professional cards with much larger memory capacities.
Meta says the compression process does not produce a noticeable performance loss.
DFlash is used to speed up token generation
The second important part of Muse Glimmer is its DFlash drafter model.
Instead of forcing the main model to generate every token sequentially, DFlash predicts larger chunks of text in advance.
Muse Glimmer then checks those predictions.
Correct parts are kept, while incorrect parts are rejected and regenerated.
The idea is that verifying predicted text can be faster than generating every token from the beginning.
Meta says this technique produces a 3.1 times response speed improvement on an RTX 5090.
The same approach reportedly improves performance by 1.8 times on an Apple M5 Max and 1.5 times on an M4 Max.
Meta is returning to open weight AI
Muse Glimmer also marks a renewed push by Meta in open weight models.
The launch arrives while Chinese open models have been gaining substantial usage.
According to the figures in the report, global AI model activity reached 69 trillion tokens during the week beginning August 3, 2026.
Chinese models accounted for 34.25 trillion of those tokens, while US models contributed 9.17 trillion.
That week represented the fifteenth consecutive week in which Chinese models led the measured token count.
DeepSeek V4 Flash was one of the fastest growing models in the dataset, recording a 570 percent week over week increase.
Open model competition is becoming more important
Muse Glimmer gives Meta another entry in a market that has become increasingly competitive.
The model is designed around local execution rather than requiring large cloud infrastructure, which could make it useful for developers who want private inference, local coding assistance and always available agents.
Its memory requirements are especially important.
A 30 billion parameter model that fits on a single 24GB or 32GB GPU is much easier to deploy than a model requiring several high end accelerators.
The DFlash approach also addresses another important limitation of local models by improving response speed without increasing the main model size.
Muse Glimmer therefore combines three priorities that are becoming central to local AI: lower memory use, faster token generation and open weight access.
Whether it changes broader open model usage remains uncertain, but it gives Meta a more competitive option at a time when Chinese models are seeing strong global adoption.
Discussion (0)
Be the first to comment.