Tied embeddings (= output head) are kept at 8-bit. LiquidAI/LFM2.5-1.2B-JP-202606 ties embed_tokens with the LM head, so a uniform low-bit quant would degrade both the input lookup and the output logits.
The MoE router gate is kept in fp32 — not applicable here, since this is a dense model with no experts.

Use with mlx-lm

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("LiquidAI/LFM2.5-1.2B-JP-202606-MLX-8bit")
messages = [{"role": "user", "content": "日本の首都はどこですか？"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt, max_tokens=128, verbose=True))

Conversion

Exported with liquidmlx. See the base model card for license, training, and intended-use details.