Nova-LFM 1.2B is a state-of-the-art "Thinking" model engineered for efficiency. It brings reasoning capabilities—typically reserved for 7B+ models—down to the 1.2B parameter class, making it possible to run complex logic chains on edge devices like Raspberry Pis, older smartphones, and laptops.
Built on the Liquid LFM-2.5 architecture, this model utilizes a novel Hybrid Self-Correction fine-tuning method. It pauses to "think" (denoted by <think> tags), verifies its own logic, and corrects errors before generating a final answer.
🌟 Key Capabilities
System 2 Thinking: Breaks down multi-step math and logic problems instead of guessing.
Edge-Native: Runs on <3GB VRAM (FP16) or <1GB (Quantized).
Balanced Profile: engineered to excel at Math (GSM8K) without sacrificing General Knowledge (MMLU).
📊 Benchmark Performance
Nova-LFM outperforms the industry standard (Llama 3.2 1B) and larger models (Gemma 2 2B) in mathematical reasoning, while maintaining a higher general knowledge score than specialized "math-only" models.
Benchmark Comparison
Model
Parameters
Math Reasoning (GSM8K)
Knowledge (MMLU)
Verdict
Nova-LFM (Ours)
1.2B
53.5% 🚀
50.1%
Best Balance
Llama 3.2 Instruct
1.0B
44.4%
42.9%
Baseline
Gemma 2
2.6B
46.4%
51.7%
Inefficient
DeepSeek R1 Distill
1.5B
69.9%
39.2% 🔻
Knowledge Collapse
SmolLM2
1.7B
31.1%
48.9%
Weak Reasoning
Note: Scores represent 5-shot evaluations using EleutherAI LM Harness. DeepSeek R1 shows significant degradation in general knowledge (MMLU < 40%) despite high math scores. Nova-LFM maintains >50% MMLU for general-purpose usability.
🚀 Quick Start
Option 1: Python (Transformers)
Requires transformers >= 4.46.0
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
2import torch
34model_id ="NovachronoAI/Nova-LFM-1.2B-Thinking"56# Load the model7tokenizer = AutoTokenizer.from_pretrained(model_id)8model = AutoModelForCausalLM.from_pretrained(9 model_id,10 torch_dtype=torch.float16,11 device_map="auto"12)1314# Define the prompt15prompt ="""Below is an instruction that describes a task. Write a response that appropriately completes the request.
1617### Instruction:
18If I have 3 apples and eat one, then buy two more, how many do I have?
1920### Response:
21"""2223inputs = tokenizer(prompt, return_tensors="pt").to("cuda")2425# Generate with reasoning (Temperature 0.6 recommended)26outputs = model.generate(27**inputs,28 max_new_tokens=512,29 temperature=0.6,30 do_sample=True31)3233print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Option 2: Ollama (Local)
For the fastest inference on CPU/Mac/Edge:
ollama run hf.co/NovachronoAI/Nova-LFM-1.2B-Thinking-GGUF
🔬 Methodology: Hybrid Self-Correction
Standard small models often hallucinate because they rush to predict the next token. Nova-LFM was trained to pause and verify.
Dataset Construction: We curated a hybrid dataset combining:
11k Standard CoT: High-quality linear reasoning chains (Step A → Step B).
4k Self-Correction Traces: Synthetic data where the model explicitly doubts itself (e.g., "Wait, that calculation seems off..."), catches the error, and corrects it.
Fine-Tuning: Trained using Unsloth with LoRA adapters targeting the unique Liquid Neural Network layers (in_proj, out_proj, w1-w3).
This dual approach teaches the model that backtracking is allowed, significantly reducing logic errors in multi-step tasks.
⚠️ Limitations
Hallucination: As a 1.2B model, it does not possess the vast world knowledge of a 70B model. It may hallucinate obscure facts or dates.
Token Artifacts: Rarely, raw training tags like [Reasoning] may appear in the output.
Context: Optimized for short-to-medium reasoning tasks (up to 8k context).
📜 Citation
If you use this model in your research or application, please cite:
bibtex
1@misc{nova-lfm-2026,
2 title = {Nova-LFM: Scalable System 2 Reasoning at the 1B Scale},
3 author = {Novachrono},
4 year = {2026},
5 publisher = {HuggingFace},
6 url = {[https://huggingface.co/NovachronoAI/Nova-LFM-1.2B-Thinking](https://huggingface.co/NovachronoAI/Nova-LFM-1.2B-Thinking)}
7}
Built with ❤️ by NovachronoAI using Unsloth & Liquid AI
