LLM EXPLORER 59,420 MODELS INDEXED

Ahma 3B by Finnish-NLP

By Finnish-NLP · 951 downloads
HF Repo ↗ Copy URL

Ahma 3B is an open-source language model by Finnish-NLP. Features: 3b LLM, VRAM: 7.3GB, Context: 2K, License: apache-2.0, LLM Explorer Score: 0.12.

  Arxiv:2302.06675   Arxiv:2302.13971   Arxiv:2305.16264   Conversational Dataset:finnish-nlp/culturax f... Dataset:finnish-nlp/hplt 1.2 f... Dataset:finnish-nlp/reddit fi ... Dataset:finnish-nlp/wikipedia ... Dataset:intfloat/multilingual ...   Doi:10.57967/hf/4011   Fi   Finnish   Llama   Region:us   Safetensors   Sharded   Tensorflow
Model Card on HF 🤗: https://huggingface.co/Finnish-NLP/Ahma-3B 

Ahma 3B Parameters and Internals

Model Type 
text generation
Use Cases 
Areas:
Research, Commercial applications
Primary Use Cases:
Text generation, Fine-tuning for downstream tasks
Limitations:
Not suitable for multilingual use, Biased predictions due to training data characteristics, Cannot generate code
Considerations:
Can benefit from fine-tuning for instruction following
Additional Notes 
Utilizes 2-stage pretraining approach to boost instruction-following capabilities.
Supported Languages 
fi (native)
Training Details 
Data Sources:
Finnish-NLP/CulturaX_fi_cleaned, Finnish-NLP/HPLT_1.2_fi_cleaned, Finnish-NLP/wikipedia_20231101_fi_cleaned, Finnish-NLP/Reddit_fi_2006_2022, Yle Finnish News Archive 2011-2018, Yle Finnish News Archive 2019-2020, Finnish News Agency Archive (STT), The Suomi24 Sentences Corpus, Project Lönnrot, Finnish parliament speeches, multilingual_cc_news, fi-news-corpus, Finnish higher education public theses, Finnish single-turn instruction-following datasets
Data Volume:
23 billion words
Methodology:
Pretrained on Finnish language, resampling and filtering techniques used, included instruction-following examples mixed in.
Context Length:
2048
Training Time:
N/A
Hardware Used:
TPUv4-32
Model Architecture:
3B parameter, decoder-only transformer
Input Output 
Input Format:
Tokenized text input as structured prompts with context.
Accepted Modalities:
text
Output Format:
Generated text
Performance Tips:
Utilize optimal generation settings like repetition penalty, and proper prompt formatting.
Release Notes 
Version:
N/A
Date:
N/A
Notes:
Model checkpoints were pushed every 100,000 training steps.
LLM NameAhma 3B
Repository 🤗https://huggingface.co/Finnish-NLP/Ahma-3B 
Model Size3b
Required VRAM7.3 GB
Updated2026-08-08
MaintainerFinnish-NLP
Model Typellama
Model Files  5.0 GB: 1-of-2   2.3 GB: 2-of-2
Supported Languagesfi
Model ArchitectureLlamaForCausalLM
Licenseapache-2.0
Context Length2048
Model Max Length2048
Transformers Version4.42.0.dev0
Tokenizer ClassPreTrainedTokenizerFast
Vocabulary Size64256
Torch Data Typefloat16

Best Alternatives to Ahma 3B

Best Alternatives
Context / RAM
Downloads
Likes
Nanbeige4.1 3B256K / 7.9 GB208751161
Nanbeige4.1 3B Heretic256K / 7.9 GB14950
Nanbeige4.1 3B Heretic Mxfp4256K / 7.9 GB342
Nanbeige4.1 3B Bf16256K / 7.9 GB473
...anbeige4.1 3B Heretic Bfloat16256K / 7.9 GB331
Nanbeige4.1 3B Heretic256K / 7.9 GB180
Nanbeige4.1 3B Heretic256K / 7.9 GB141
ISA 03 Mini 3B Hybrid Preview256K / 6.5 GB44
ToolMind Web 3B192K / 7.9 GB15628
Llama 3.2 3B Instruct128K / 6.5 GB18295722368
Note: green Score (e.g. "73.2") means that the model is better than Finnish-NLP/Ahma-3B.