Muse-Spark-Open-816B-A42B-oQ4e
This model was quantized using oQ (oMLX v0.7.1) mixed-precision quantization with oQe importance-matrix quantization.
Re-uploaded on 2026-10-16 at 08:41 UTC in the layout of mlx-lm 0.32.2, so it loads in both mlx-lm 0.32.2 and oMLX. This version keeps the lm_head in BF16, which mlx-lm 0.32.2 needs for this model, and gives q, k, v and the attention output gate of each layer one bit width, as mlx-lm stores them as one projection.
A mixed-precision MLX checkpoint of meta-models/Muse-Spark-Open-816B-A42B, Meta's 816B total / 42B active Mixture-of-Experts model with a 1M-token context window, built for coding and agentic work. It includes the language model and the vision encoder. The audio encoder is not included.
The routed experts of layers 1-60 are passed through unchanged from the official NVFP4 release (meta-models/Muse-Spark-Open-816B-A42B-NVFP4), with their per-expert tensor scales.
The effective average is 4.649 bits/weight for the complete checkpoint.
Quantization and Bit Distribution
oQ4e allocates bits by measured layer sensitivity, then rounds each affine group with an importance matrix (128 samples x 512 tokens) and a weighted least-squares refit of its scale and bias. The NVFP4 experts pass through unchanged.
| Storage format | Logical weights | Share of total | Tensor storage | Effective bits/weight |
|---|---|---|---|---|
| NVFP4, group 16 | 773.094B | 94.060% | 405.000 GiB | 4.50 |
| Affine 4-bit, group 64 | 25.770B | 3.135% | 13.500 GiB | 4.50 |
| Affine 8-bit, group 64 | 19.338B | 2.353% | 19.136 GiB | 8.50 |
| BF16 | 3.585B | 0.436% | 6.678 GiB | 16.00 |
| FP32 | 0.132B | 0.016% | 0.492 GiB | 32.00 |
| Total | 821.920B | 100% | 444.806 GiB | 4.649 |
Storage Breakdown
| Component | Logical weights | Storage |
|---|---|---|
| Language backbone | 819.992B | 473.751 GB / 441.215 GiB |
| Vision encoder | 1.928B | 3.856 GB / 3.591 GiB |
| Total | 821.920B | 477.607 GB / 444.806 GiB |
[!WARNING] This checkpoint takes about 478 GB, which is more than a single 512 GB Mac can run reliably. For one 512 GB Mac, use the oQ3e quant instead.
Use with mlx-lm
pip install -U "mlx-lm>=0.32.2"
mlx_lm.generate --model mlx-community/Muse-Spark-Open-816B-A42B-oQ4e \
--prompt "Explain convolutional networks like I am 5" \
--max-tokens 5000
Use with oMLX
Download the model in oMLX and chat with it in the built-in chat UI. You can add images and set the reasoning effort from low to xhigh. Reasoning cannot be turned off, and the default (Auto) uses xhigh.
You can also use
- mlx-community/Muse-Spark-Open-816B-A42B-oQ3e: the 3-bit oQ quant.
- mlx-community/Muse-Spark-Open-816B-A42B-3bit: the standard mlx-lm 3-bit quant.
- mlx-community/Muse-Spark-Open-816B-A42B-4bit: the standard mlx-lm 4-bit quant.
License
Muse Model License, following the source model. See LICENSE.md and USE_POLICY.md.