I couldn't find any information regarding the creators of this model (GrootTeam), be sure to verify the outputs of the model and to not rely on them blindly.

Groot-19B-A3B-261023-oQ4e-fp16-mtp

This model was quantized using oQ (oMLX v0.7.3.dev2+rdsm.dflash2) mixed-precision quantization.

The effective average is 5.294 bits/weight for the complete checkpoint.

Quantization and Bit Distribution

Storage format	Logical weights	Share of total	Tensor storage	Effective bits/weight
Affine 4-bit, group 64	17.969B	92.472%	9.414 GiB	4.50
Affine 5-bit, group 64	0.356B	1.833%	0.228 GiB	5.50
Affine 6-bit, group 64	0.002B	0.011%	0.002 GiB	6.50
Affine 8-bit, group 64	0.638B	3.281%	0.631 GiB	8.50
FP16	0.020B	0.105%	0.038 GiB	16.00
FP32	0.447B	2.298%	1.664 GiB	32.00
Total	19.432B	100%	11.976 GiB	5.294

Storage Breakdown

Component	Logical weights	Storage
Language model	18.544B	10.810 GB / 10.068 GiB
Vision	0.447B	1.786 GB / 1.664 GiB
MTP head	0.442B	0.262 GB / 0.244 GiB
Total	19.432B	12.859 GB / 11.976 GiB

Benchmarks vs full model

A subset of the benchmarks was run with thinking disabled, solely to confirm that quantisation did not degrade performance, not for comparison with other models.

Benchmark	Groot-19B-A3B-261023-oQ4e-fp16-mtp	Groot-19B-A3B-261023
gsm8k	96.0% (96/100) · 73.2 s	94.0% (94/100) · 133.4 s
arc_challenge	85.0% (85/100) · 12.6 s	85.0% (85/100) · 18.3 s
mmlu	63.0% (63/100) · 40.0 s	65.0% (65/100) · 51.8 s
truthfulqa	78.0% (78/100) · 15.0 s	77.0% (77/100) · 15.3 s
mmlu_pro	61.0% (61/100) · 222.7 s	57.0% (57/100) · 416.1 s

Quick read: the -oQ4e-fp16-mtp variant is faster on every benchmark (often ~2× on gsm8k and mmlu_pro) and scores equal or better everywhere except a small mmlu drop (63.0 vs 65.0).