Model reference · open weights

xls-r-danish-mc

Audio MediaCatch Speech→text 1 build Open weights 2k dl/mo

xls-r-danish-mc is an open-weight audio or speech model from MediaCatch. xls-r-300m-danish-mc-v2 (FP32) weighs 631 MB; the smallest configuration that runs it is RTX 3060 12 GB.

Request Rent a GPU

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

xls-r-danish-mc is an automatic speech recognition model developed by MediaCatch and published as MediaCatch/xls-r-300m-danish-mc-v2.
It contains 315 million parameters and is distributed under the apache-2.0 license.
The model achieves a word error rate of 0.0875 on its evaluation set.

Summary of the MediaCatch/xls-r-300m-danish-mc-v2 model card, 2026-10-01

Hugging Face ↗

What it is

Released by	MediaCatch
Released	2025-09-29
Parameters	315M
VRAM	631 MB for the weights

What it runs on

Memory and cards for xls-r-300m-danish-mc-v2 (FP32)

631 MBweights, file size

1.6 GBruntime overhead

Card	One stream	Memory
RTX 3060 12 GB	fits	11.6 GB
RTX 4060 Ti 16 GB	fits	15.4 GB
RTX 3090 24 GB	fits	23.4 GB
RTX 4090 24 GB	fits	23.4 GB
RTX 5090 32 GB	fits	31.0 GB
L40S 48 GB	fits	44.0 GB
A100 80 GB	fits	78.2 GB
H100 80 GB	fits	78.1 GB
RTX PRO 6000 Blackwell 96 GB	fits	93.8 GB
DGX Spark (GB10) 128 GB unified	fits	107 GB
H200 141 GB	fits	138 GB
B200 180 GB	fits	176 GB

From the model card

What MediaCatch says about xls-r-danish-mc

Read the model card

This model is a fine-tuned version of MediaCatch/xls-r-300m-danish-mc on the Preprocessed Dataset dataset. It achieves the following results on the evaluation set:

Loss: 0.1570
Wer: 0.0875

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 3
eval_batch_size: 6
seed: 69
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 4
total_train_batch_size: 48
total_eval_batch_size: 24
optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 500
num_epochs: 10.0

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
No log	0	0	3.1518	0.1154
0.1233	0.0991	500	0.1684	0.1039
0.1122	0.1982	1000	0.1536	0.1059
0.109	0.2973	1500	0.1497	0.1032
0.1027	0.3965	2000	0.1471	0.1036
0.1082	0.4956	2500	0.1446	0.1053
0.105	0.5947	3000	0.1440	0.1059
0.0986	0.6938	3500	0.1436	0.1041
0.1037	0.7929	4000	0.1404	0.1015
0.0988	0.8920	4500	0.1391	0.1008
0.0979	0.9911	5000	0.1373	0.0995
0.0833	1.0902	5500	0.1385	0.0991
0.0855	1.1893	6000	0.1407	0.0990
0.0848	1.2884	6500	0.1390	0.0975
0.0845	1.3875	7000	0.1365	0.0986
0.0824	1.4866	7500	0.1373	0.0975
0.0842	1.5858	8000	0.1353	0.0979
0.0837	1.6849	8500	0.1334	0.0968
0.0823	1.7840	9000	0.1348	0.0960
0.0851	1.8831	9500	0.1331	0.0957
0.0843	1.9822	10000	0.1299	0.0949
0.0707	2.0813	10500	0.1337	0.0926
0.0709	2.1804	11000	0.1332	0.0958
0.0742	2.2795	11500	0.1316	0.0944
0.0742	2.3786	12000	0.1356	0.0959
0.0719	2.4777	12500	0.1323	0.0969
0.0739	2.5768	13000	0.1286	0.0951
0.0695	2.6760	13500	0.1315	0.0957
0.0741	2.7751	14000	0.1310	0.0940
0.0729	2.8742	14500	0.1303	0.0970
0.0695	2.9733	15000	0.1316	0.0939
0.0637	3.0724	15500	0.1353	0.0955
0.0664	3.1715	16000	0.1333	0.0940
0.0635	3.2706	16500	0.1370	0.0941
0.0652	3.3697	17000	0.1334	0.0937
0.0653	3.4688	17500	0.1320	0.0957
0.0654	3.5679	18000	0.1365	0.0938
0.0633	3.6670	18500	0.1363	0.0943
0.0642	3.7661	19000	0.1316	0.0926
0.0622	3.8653	19500	0.1282	0.0906
0.0653	3.9644	20000	0.1334	0.0904
0.0585	4.0634	20500	0.1363	0.0914
0.057	4.1625	21000	0.1334	0.0935
0.0591	4.2617	21500	0.1370	0.0914
0.0538	4.3608	22000	0.1357	0.0929
0.0586	4.4599	22500	0.1379	0.0916
0.0556	4.5590	23000	0.1378	0.0925
0.0574	4.6581	23500	0.1353	0.0898
0.0545	4.7572	24000	0.1371	0.0912
0.0572	4.8563	24500	0.1320	0.0895
0.0546	4.9554	25000	0.1361	0.0908
0.0485	5.0545	25500	0.1429	0.0926
0.054	5.1536	26000	0.1401	0.0912
0.0507	5.2527	26500	0.1406	0.0888
0.0519	5.3519	27000	0.1416	0.0902
0.0524	5.4510	27500	0.1403	0.0903
0.05	5.5501	28000	0.1395	0.0890
0.0503	5.6492	28500	0.1439	0.0892
0.0528	5.7483	29000	0.1402	0.0905
0.0503	5.8474	29500	0.1424	0.0902
0.051	5.9465	30000	0.1412	0.0890
0.0471	6.0456	30500	0.1447	0.0893
0.0461	6.1447	31000	0.1511	0.0885
0.0436	6.2438	31500	0.1505	0.0898
0.0483	6.3429	32000	0.1458	0.0884
0.0457	6.4420	32500	0.1449	0.0886
0.0465	6.5412	33000	0.1430	0.0880
0.0449	6.6403	33500	0.1487	0.0892
0.0455	6.7394	34000	0.1491	0.0883
0.0483	6.8385	34500	0.1476	0.0884
0.0485	6.9376	35000	0.1449	0.0885
0.0445	7.0367	35500	0.1504	0.0878
0.0429	7.1358	36000	0.1544	0.0887
0.0429	7.2349	36500	0.1507	0.0885
0.0449	7.3340	37000	0.1499	0.0890
0.0414	7.4331	37500	0.1522	0.0878
0.0414	7.5322	38000	0.1519	0.0888
0.0405	7.6313	38500	0.1540	0.0878
0.0424	7.7305	39000	0.1535	0.0884
0.042

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

Details

Type	Audio & music
Task	Speech→text
Runs with	transformers
Based on	MediaCatch/xls-r-300m-danish-mc
Popularity	2k downloads / month
Weights	631 MB, xls-r-300m-danish-mc-v2 (FP32, file size)