Muse Spark Open
Run with https://llama.app
llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q2_K
Source models
Files
| Quant | Size | Notes |
|---|---|---|
| BF16 | 1.64 TB | Original precision |
| Q8_0 | 872 GB | Near-lossless. All weight matrices at Q8_0 |
| Q4_K_S | 470 GB | 4-bit quant |
| Q2_K | 304 GB | imatrix-calibrated. Fits in 512 GB of unified memory |
| mmproj (BF16, F16, Q8_0) | 4.1 / 4.1 / 2.2 GB | Vision and audio encoders |
Notes
- Muse Spark Open is an 816B-parameter mixture-of-experts model (256 experts, ~42B active per token) with image and audio input.
- The Q2_K output quantizes the routed expert gate/up projections to Q2_K and the expert down projections to Q3_K. The per-layer latent projections (
ffn_latent_down,ffn_latent_up) and the attention output gate (attn_gate) are kept at Q8_0. - The Q2_K is calibrated with an imatrix computed from the Q8_0 on bartowski's v6 calibration data.
- Reasoning depth is set with
reasoning_effort:low,medium,highorxhigh(the default). A system prompt containing aReasoning strength: N.line overridesreasoning_effort. - Recommended sampling (stored in the GGUF metadata):
temperature 1.0,top_p 1.0,top_k 64.