from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler, make_logits_processors

model, tokenizer = load("LiquidAI/LFM2.5-1.2B-Thinking-4bit")

prompt = "solve 2+2*2"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

sampler = make_sampler(temp=0.1, top_k=50, top_p=0.1)
logits_processors = make_logits_processors(repetition_penalty=1.05)

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

Chat Format

The model uses the ChatML format:

<|startoftext|><|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant
{assistant_response}<|im_end|>

Thinking Format

The model outputs chain-of-thought reasoning inside <think> tags before the final answer:

<think>
[Internal reasoning process...]
</think>

[Final answer]

Example

Prompt:

<|startoftext|><|im_start|>user
solve 2+2*2<|im_end|>
<|im_start|>assistant

Model output:

<think> Okay, let me try to solve this problem: "2+2*2". Hmm, first I need to make sure I understand what exactly is being asked here. The expression is 2 + 2*2. Alright, so maybe it's just a simple arithmetic problem. Let me break it down step by step.

First, looking at the expression 2 + 2*2. I know in math, the order of operations is important here. Multiplication comes before addition, right? So the 2*2 would be calculated first. So 2 times 2 is 4. Then add that to the initial 2. So 2 + 4 equals 6.
</think>

The answer is **6**.

Following the order of operations (PEMDAS/BODMAS), multiplication is performed before addition:
- First: 2 × 2 = 4
- Then: 2 + 4 = 6

Stripping Thinking from History

The chat template automatically strips <think> content from previous assistant messages in multi-turn conversations. To preserve thinking in history, use:

python

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    keep_past_thinking=True  # Preserve thinking in history
)

Tool Calling

The model supports function calling with a specific format.

Tool Definition

Tools are defined as JSON in the system prompt:

List of tools: [{"name": "tool_name", "description": "...", "parameters": {...}}]

Tool Call Format

The model generates tool calls using special tokens:

<|tool_call_start|>[function_name(arg1="value1", arg2="value2")]<|tool_call_end|>

Tool Response Format

Tool results are provided in a tool role message:

<|im_start|>tool
[{"result": "..."}]<|im_end|>

License

This model is released under the LFM 1.0 License.

LiquidAI/LFM2.5-1.2B-Thinking-MLX-4bit

LFM2.5-1.2B-Thinking-4bit

Model Details

Recommended Sampling Parameters

Use with mlx