LLM EXPLORER 62,323 MODELS INDEXED

Facebook Opt 125M Qcqa Ub 6 Best For KV Cache by xformAI

By xformAI · 58 downloads

Facebook Opt 125M Qcqa Ub 6 Best For KV Cache is an open-source language model by xformAI. Features: 125m LLM, VRAM: 0.5GB, Context: 2K, License: mit, LLM Explorer Score: 0.15, Arc: 24.2, HellaSwag: 25, MMLU: 23.1.

  En   Endpoints compatible   Opt   Pytorch   Region:us

Facebook Opt 125M Qcqa Ub 6 Best For KV Cache Parameters and Internals

Additional Notes 
This implementation is supposed to be more efficient than the corresponding GQA one.
Supported Languages 
en (native)
Training Details 
Model Architecture:
The original MHA architecture is preserved, but instead of having a single K/V head, different K/V heads corresponding to the same group have the same mean-pooled K or V values with up to 6 groups of KV heads per layer instead of original 12 KV heads in the MHA implementation.
LLM NameFacebook Opt 125M Qcqa Ub 6 Best For KV Cache
Repository πŸ€—https://huggingface.co/xformAI/facebook-opt-125m-qcqa-ub-6-best-for-KV-cache 
Model Size125m
Required VRAM0.5 GB
Updated2026-08-01
MaintainerxformAI
Model Typeopt
Model Files  0.5 GB
Supported Languagesen
Model ArchitectureOPTForCausalLM
Licensemit
Context Length2048
Model Max Length2048
Transformers Version4.26.1
Tokenizer ClassGPT2Tokenizer
Beginning of Sentence Token</s>
End of Sentence Token</s>
Unk Token</s>
Vocabulary Size50272
Torch Data Typefloat32
Activation Functionrelu
Errorsreplace

Best Alternatives to Facebook Opt 125M Qcqa Ub 6 Best For KV Cache

Best Alternatives
Context / RAM
Downloads
Likes
Opt 125M2K / 0.3 GB9414291252
...25M Gqa Ub 6 Best For KV Cache2K / 0.5 GB2070
...125M Qcqa Ub 6 Best For Q Loss2K / 0.5 GB570
Galactica 125M Cot2K / 0.5 GB50
Galactica Ref2K / 0.5 GB50
Galactica 125M DPO2K / 0.5 GB70
Galactica 125M DPO Pos2K / 0.5 GB50
BertQA2K / 0.5 GB50
BertQA2K / 0.5 GB50
Opt 125M Quantized Brevitas2K /  GB60
Note: green Score (e.g. "73.2") means that the model is better than xformAI/facebook-opt-125m-qcqa-ub-6-best-for-KV-cache.