[!Important] Model trained on Claude Sonnet 4.6 Adaptive Thinking (The claude.ai Claude, not the API Claude). The training data is PRIVATE.

LFM2.5-1.2B-Distilled-Claude-4.6 (Liquid Claude)

LFM2.5-1.2B-Distilled-Claude-4.6 (Liquid Claude) is a distillation of Claude into LFM2.5-1.2B-Thinking via LoRA.

Training data info:

THINK BLOCK PATTERNS Agentic think blocks (Action/Observation): 8473 Pure reasoning think blocks: 3005

CONSECUTIVE USER MESSAGES Total consecutive user-user pairs: 319

MESSAGE LENGTH STATS (chars)

Assistant messages total: 11738 With agentic tool calls in think: 2151 (18.3%)

Total chars in dataset: 60,487,999 Approx tokens (~4 chars/token): 15,121,999

Conversations with <=2 messages (system+1): 121 Conversations with >5 think blocks in a single assistant msg: 319

Use model

from transformers import pipeline

pipe = pipeline("text-generation", model="FlameF0X/LFM2.5-1.2B-Distilled-Claude-4.6")
messages = [
    {"role": "system", "content": "You are a helpful assistant."}, # I RECOMMEND TO KEEP THIS FOR STABILITY! But you can change the system.
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

Sample chat:

(Ignore the fact that it took 1min to reason, i got a i3-6006u / 12GB as hardware and running the f16 quantization)

Benchmark