Jugurtha-4B-MLX-4bit (MLX)

4-bit quantized MLX format for Jugurtha-4B, the first foundation model for Kabyle (Taqbaylit, kab, Latin script) with native chain-of-thought reasoning.

Optimized natively for Apple Silicon (M1, M2, M3, M4) using Apple's MLX framework. Runs with high token throughput and minimal unified memory footprint (~2.5 GB RAM).

---

Quick Start

Installation

pip install mlx-lm

Python API

from mlx_lm import generate, load

model, tokenizer = load("agbalu/Jugurtha-4B-MLX-4bit")

messages = [{"role": "user", "content": "Segzi-yi-d s wegzul d acu i d Agbalu."}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=512,
    verbose=True,
)
print(response)

Command Line Interface

# Direct generation
mlx_lm.generate \
  --model agbalu/Jugurtha-4B-MLX-4bit \
  --prompt "<|im_start|>user\nAzul, amek telliḍ?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n" \
  --max-tokens 512

# Interactive local server / web chat
mlx_lm.server --model agbalu/Jugurtha-4B-MLX-4bit

---

Native Reasoning & Thinking

Jugurtha-4B supports native dual-mode reasoning on Apple Silicon:

• Thinking Mode ON: Start assistant generation with <think>\n to let the model articulate its internal thoughts in Taqbaylit before delivering the final answer.
• Thinking Mode OFF: Start assistant generation with <think>\n\n</think>\n\n for immediate direct answers.

---

Full Model & Benchmark Details

For complete evaluations (FLORES+, KabBench, IFEval, GSM8K), training details, and safetensors weights, visit the primary repository: agbalu/Jugurtha-4B.

Licence

Apache-2.0