Zera-24B-4bit

Zera-24B-4bit

Zera is a 24B-parameter, 4-bit MLX language model fine-tuned for coding, debugging, technical explanations, general conversation, grammar, and vocabulary. It is a standalone fused model: users do not need a separate adapter.

Use with MLX

pip install mlx-lm

from mlx_lm import generate, load

model, tokenizer = load("Zexiry/Zera-24B-4bit")
messages = [{"role": "user", "content": "Introduce yourself, then write a Python trie."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=800, verbose=False))

Zera's default chat template supplies its assistant identity when an application does not provide a system prompt. Applications can still provide their own system prompt normally.

Technical details

• Parameters: 23.57B
• Weight format: MLX 4-bit, group size 64
• Fine-tuning: QLoRA, rank 32, 24 adapted layers
• License: Apache-2.0

The base_model metadata above is retained for reproducibility, attribution, and license compliance. In conversation, the assistant identity is Zera.