Lucasgitguy-co/K2-Horizon-0.9B-System1-MLX-8bit

K2-Horizon-0.9B-System1 (MLX 8-bit)

This is the 8-bit MLX build of K2-Horizon-0.9B-System1. That model is a no-thinking ("System 1") fine-tune of IFM/K2-Horizon-0.9B: it answers right away instead of writing a long think block first. Results, method and limitations are on the main card.

K2-Horizon-0.9B is Llama-shaped, so this folder uses mlx-lm's llama model type, with YaRN RoPE ×16 from an 8K original context.

pip install mlx-lm
mlx_lm.generate --model Lucasgitguy-co/K2-Horizon-0.9B-System1-MLX-8bit --prompt "What is 17*23?"

from mlx_lm import load, generate
model, tok = load("Lucasgitguy-co/K2-Horizon-0.9B-System1-MLX-8bit")
p = tok.apply_chat_template([{"role": "user", "content": "Give me one tip for sleeping better."}],
                            add_generation_prompt=True, tokenize=False)
print(generate(model, tok, p, max_tokens=256))
# reasoning_effort="high" in apply_chat_template switches to full thinking

Checked: greedy answers to 5 test prompts match the bf16 model (4-bit drifts slightly in wording). Apache-2.0.