Mistral-7B-v0.1 config.json: sliding_window 4096, GQA 8 kv heads

Fetched 2026-10-18. config.json fields used: hidden_size 4096, num_hidden_layers 32, num_attention_heads 32, num_key_value_heads 8, max_position_embeddings 32768, sliding_window 4096, torch_dtype bfloat16, vocab_size 32000. head_dim is hidden_size/num_attention_heads = 128 and is not a separate field. This is the base Mistral-7B-v0.1 checkpoint, not the OpenHermes finetune already in models/.