mlx-community/LFM2.5-1.2B-Thinking-OptiQ-4bit

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. All OptiQ quants · Docs

An OptiQ mixed-precision quant of LiquidAI/LFM2.5-1.2B-Thinking. 820 MB on disk, down from 2.34 GB at bf16, and it scores 84% on IFEval and 83% on GSM8K at that size.

This is the reasoning variant of LFM2.5-1.2B. It thinks before answering, which is worth a lot at this size: against the Instruct model it gains 21 points on MMLU and 13 on GSM8K.

LFM2.5 is a hybrid architecture: convolutional blocks interleaved with full attention, so only a few blocks carry a KV cache. OptiQ measures each layer's sensitivity and assigns per-layer bit-widths, keeping 8-bit where it matters and 4-bit elsewhere.

What it is

Property	Value
Base	LiquidAI/LFM2.5-1.2B-Thinking
Architecture	lfm2 — hybrid conv + full attention
Method	OptiQ mixed-precision, sensitivity-driven (bf16 reference)
On disk	820 MB (bf16: 2.34 GB)
Context	128k
KV cache	mixed-precision kv_config.json bundled (6 attention layers, 4.67 avg bits)

Capability Score

Six-metric mean, the standard OptiQ eval. Run with reasoning enabled.

Metric	Score
MMLU (5-shot, 969 samples)	64.7%
GSM8K (1000 samples)	82.8%
IFEval (full set, strict)	84.3%
BFCL-V3 simple (200 calls)	45.5%
HumanEval (164 problems, pass@1)	47.6%
HashHop (long-context retrieval)	0.0%
Capability Score (mean of 6)	54.14

Knowledge and instruction-following are the strengths, and 83% on GSM8K is unusual for a model under a gigabyte. Long-context multi-hop retrieval is the weak spot: at this size the model cannot produce the hash format the task requires, so it scores zero rather than partially.

Run it

pip install mlx-optiq

from mlx_lm import load, generate

model, tok = load("mlx-community/LFM2.5-1.2B-Thinking-OptiQ-4bit")
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "What is 17 * 23? Think briefly then answer."}],
    tokenize=False, add_generation_prompt=True,
)
print(generate(model, tok, prompt=prompt, max_tokens=2048))

Give it room. The model spends its first few hundred tokens thinking, and that budget comes out of max_tokens. Cap it too low and the whole allowance is consumed inside the reasoning block, so you get an empty answer back rather than a short one. 2048 is a sensible floor for anything non-trivial.

Serve it with the bundled mixed-precision KV cache:

optiq serve --model mlx-community/LFM2.5-1.2B-Thinking-OptiQ-4bit --kv-config kv_config.json

Tool calling

LFM2.5 writes Pythonic calls between special tokens:

<|tool_call_start|>[get_weather(city="Paris")]<|tool_call_end|>

Pass tools= to apply_chat_template and they are rendered into the system prompt.

Links

• Project website: mlx-optiq.com
• All OptiQ quants: mlx-optiq.com/models
• Base model: LiquidAI/LFM2.5-1.2B-Thinking