Yi000153/Zera-35B-A3B-4bit

Zera is a standalone 4-bit MLX language model fine-tuned for natural conversation, programming, instruction following, and a consistent Zera identity. The QLoRA adapter has been fused into the model weights, so no separate adapter is required.

• Parameters: 34.66B total, approximately 3B active per token
• Format: MLX, 4-bit quantized
• Base: mlx-community/Qwen3.6-35B-A3B-4bit
• Held-out answer loss: 0.571
• Held-out perplexity: 1.769

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Yi000153/Zera-35B-A3B-4bit")

prompt = "Hello, who are you?"

if tokenizer.chat_template is not None:
    messages = [
        {"role": "system", "content": "You are Zera, a capable and accurate AI assistant."},
        {"role": "user", "content": prompt},
    ]
    prompt = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=False,
        enable_thinking=False,
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)

Training data

The curriculum combines hand-authored identity and general-assistant examples, preferred responses from argilla/ultrafeedback-binarized-preferences-cleaned (MIT), and test-verified programming examples from nvidia/OpenCodeInstruct (CC-BY-4.0).

Limitations

Generated answers can still be incorrect. Verify important factual claims and review generated code before using it in production.