meta-models
Muse-Spark-Open-816B-A42B
A 820B mixture-of-experts (about 42B active) model by meta-models that uses tools such as web search, can think before answering and reads images.
Tools Thinking Vision- Downloads
- 1,981 downloads last month
- Likes
- 10 likes
- Uploaded
- Uploaded 14 Oct 2026
- Parameters
- 822,036,529,664 (822B)
- Context length
- 1,048,576 tokens
- Architecture
- muse-spark
- Task
- image-text-to-text
- Licence
- muse-model-license
- Released
- 15 Sept 2026
- Last updated
- 15 Oct 2026
- Access
- Gated: accept the terms on Hugging Face
This is ggml-org's upload of meta-models/Muse-Spark-Open-816B-A42B.
Read the licence: muse-model-license
Languages listed in the model card
- English
Tools
Uses tools, so web search and the browser work best with it.
Thinking
Can think step by step before it answers.
Vision
Reads images: photos, screenshots, scanned pages.
Each version is the same model stored at a different precision: smaller files use less memory and lose a little quality.
| Version | File size | File | SHA-256 |
|---|---|---|---|
| Q2_K | 304.08 GB | Muse-Spark-Open-816B-A42B-Q2_K-00001-of-00022.gguf+21 parts | Per part |
| Q4_K_S | 470.07 GB | Muse-Spark-Open-816B-A42B-Q4_K_S-00001-of-00022.gguf+21 parts | Per part |
| Q8_0 | 871.65 GB | Muse-Spark-Open-816B-A42B-Q8_0-00001-of-00010.gguf+9 parts | Per part |
| BF16 | 1640.26 GB | Muse-Spark-Open-816B-A42B-BF16-00001-of-00022.gguf+21 parts | Per part |
Vision add-on (needed to read images): mmproj-Muse-Spark-Open-816B-A42B-Q8_0.gguf, 2.18 GB
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.
More from meta-models
On Hugging Face
Everything on this page comes from these pages and is refreshed every few hours; Hugging Face always has the latest.