Published capacity tiers
Start with the official target, then budget the runtime
Wide table: swipe or use Shift + mouse wheel to inspect every column.
| Official target | Main artifact | Exact bytes | Binary size | Component note |
|---|---|---|---|---|
| 24 GiB | K-Quant 17 GB | 16,756,681,056 | 15.606 GiB | Vision projector and DFlash drafter are separate files when selected. |
| 32 GiB | K-Quant Dynamic | 19,653,957,984 | 18.304 GiB | Vision projector and DFlash drafter are separate files when selected. |
| 64 GiB | BF16 weights | 59,553,435,272 | 55.463 GiB | The perception encoder is included; DFlash is not an optional BF16 companion here. |
Separate GGUF artifacts
Add only the companions your workload needs
mmproj-kquant.gguf
1,400,328,928 bytes · 1.304 GiB
dflash-kquant.gguf
1,631,205,312 bytes · 1.519 GiB
This is the model configuration limit, not proof that the full window fits or performs well on a particular machine.
Context memory boundary
Use an active and a conservative KV estimate
At the configured context ceiling, the architecture-derived active estimate is1.7 GiB. It counts all13 global layers across the selected context and caps the39 local layers at the2,048-token sliding window.
A conservative full-allocation boundary is6.5 GiB. It assumes all52 layers allocate the full context with16-bit KV, 2 KV heads, and head dimension128.
Planning reserve is a site assumption
This wiki adds 1.5 GiB as a small planning reserve. Meta does not publish it as a runtime requirement. Drivers, buffers, batch size, offload, cache dtype, and operating-system use can move the real peak in either direction.
Limitations
- Published target capacity is not a minimum, certification, or guarantee for any GPU, Mac, or CPU-only system.
- Exact file size is only one part of memory use; weights, KV cache, runtime buffers, and system usage coexist.
- The configured context ceiling does not promise acceptable latency or a successful full-window allocation.
- No first-party per-device peak-memory matrix was available at the last review date.
Choose the next planning step
- Use the hardware fit estimator
Compare your available memory with active and conservative planning totals.
- Choose an official GGUF
Match the main file and optional companions to your intended workload.
- Check llama.cpp support
Verify that the build includes the dated Muse Glimmer merge.