leonsarmiento/TimeCapsuleLLM-v2-llama-1.2B-8bit-mlx

This model was converted to MLX format from haykgrigorian/TimeCapsuleLLM-v2-llama-1.2B using 8-bit uniform quantization optimized for Apple Silicon.

TimeCapsuleLLM-v2-llama-1.2B is a 1.2B-parameter Llama-architecture model trained from scratch on 112 GB of historical English texts from London (1800–1875). It features 22 layers, Grouped Query Attention (GQA) with 16 query / 8 KV heads, and a custom BPE tokenizer (32K vocab). The model generates text in the style of historical English from that era.

Use with mlx

pip install -U mlx-lm

python -m mlx_lm.generate --model leonsarmiento/TimeCapsuleLLM-v2-llama-1.2B-8bit-mlx --max-tokens 256 --temperature 0.5 --prompt "The streets of London were"

Quantization Details

Metric	Value
Quantization type	8-bit uniform
Average	8.501 bits per weight
Group size	64
Method	mlx_lm
Total output size	~1.2 GB (1 shard)

Recommended Inference Parameters

This is a base completion model — it performs raw text continuation. In LM Studio, use completion mode or select the Llama 2 template for best results.

Parameter	Value
temperature	0.7
top_p	0.9
top_k	64
min_p	0.01
repetition_penalty	1.1
max_tokens	200+