LFM2.5-8B-A1B-MXFP8

MLX MXFP8 conversion of LiquidAI/LFM2.5-8B-A1B, built for Apple Silicon inference.

This is the higher-precision MXFP sibling of the MXFP4 bundle. It preserves the original Liquid chat template in chat_template.jinja.

Format

• Quantization: MLX MXFP8
• Converter output: 8.250 bits per weight
• Quantization config: mode=mxfp8, bits=8, group_size=32
• Router/gate tensors: preserved at 8-bit groups where emitted by MLX
• Local size before upload: 8.1G
• Source model: LiquidAI/LFM2.5-8B-A1B

Runtime

Use an MLX runtime with LFM2/LFM2.5 support.

from mlx_lm import load, generate

model, tokenizer = load("OsaurusAI/LFM2.5-8B-A1B-MXFP8")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is 2+2? Answer briefly."}],
    add_generation_prompt=True,
    tokenize=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=64, verbose=True))

Chat Template And Reasoning

The bundled chat_template.jinja uses Liquid's ChatML-like format:

• User and assistant turns use <|im_start|> / <|im_end|>.
• The generation prompt ends at <|im_start|>assistant\n; it does not pre-open <think>.
• Assistant reasoning may appear inside <think>...</think>.
• Tool calls use Liquid's Python-call list format inside <|tool_call_start|> and <|tool_call_end|>.

Do not force an extra synthetic <think> prefix at runtime. Let the template and model handle reasoning normally.

Verification

Local smoke run on the converted bundle:

• Prompt: What is 2+2? Answer briefly.
• Result: generated reasoning identified 4
• Reported generation speed: about 196 tok/s on a 64-token run
• Peak memory reported by the smoke run: about 8.767 GB

This is a smoke test, not a benchmark suite or accuracy evaluation.

Korean

이 모델은 LiquidAI/LFM2.5-8B-A1B를 Apple Silicon용 MLX MXFP8 형식으로 변환한 버전입니다. MXFP4보다 더 큰 고정밀 형식이며, chat_template.jinja의 기본 템플릿을 그대로 사용합니다.