Ex0bit/lfm-Nanotron

# lfm-Nanotron: 2.6B-PRISM-SFT-GSPO-AutoRoundV2

** LFM Architeture model SFT + GSPO RL + PRISM **

Model Description

lfm-Nanotron: Limited Edition 2.6B PRISM Model Access. Unlock a cutting-edge Nano sized AI model!

This is lfm-Nanotron — A Nano Sized 2.6B parameter hybrid architecture language model fine-tuned with advanced techniques you won't find in mainstream releases:

• SFT (Test-Time Supervised-Fine-Tuning) — Adaptive optimization at inference
• GSPO (Group Sequence Policy Optimization) — RL Enhanced reasoning, Instruction following, thinking, tool calling & logic
• PRISM (Projected Refusal Isolation via Subspace Modification) — State-of-the-art over-refusal/propaganda removal from LLMs
• 128K Context Window — Handle massive prompts with ease
• Agentic Tool Calling — Built for multi-turn, thinking, and instruction-following tasks

Architecture Details

Parameter	Value
Parameters	~2.6B
Hidden Size	2048
Layers	30 (22 Conv + 8 Full Attention)
Attention Heads	32
KV Heads	8 (GQA)
Vocabulary	65,536
Max Context	128,000 tokens
Architecture	Hybrid Conv + Attention (LFM2)

Available Quantizations

File	Quantization	Size	Use Case
lfm2-nanotron-ttft-gspo-prism-bf16.gguf	BF16	~4.8GB	Full precision, best quality
lfm2-nanotron-ttft-gspo-prism-Q4_K_M.gguf (+W4A16)	Q4_K_M	~1.5GB	Balanced quality/size
lfm2-nanotron-ttft-gspo-prism-Q2_K.gguf	Q2_K (+W2A16)	~0.9GB	Maximum compression

Usage

With llama.cpp

./llama-cli -m lfm2-nanotron-ttft-gspo-prism-Q4_K_M.gguf -p "Your prompt here" --temp 0.3 --min-p 0.15 --repeat-penalty 1.05

Recommended Generation Parameters

{
  "temperature": 0.3,
  "min_p": 0.15,
  "repeat_penalty": 1.05
}

Citation

If you use this model in your research, please cite:

@misc{lfm2-nanotron-2026,
  title={lfm2-Nanotron: Test-Time Fine-Tuned LFM2 with GSPO+PRISM},
  author={Exobit (Eric Elbaz)},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Ex0bit/lfm2-Nanotron}
}

License

This model is released under a custom research license. See LICENSE.md for details.

Acknowledgments

• @mlabonne & @liquidai for the LFM2 architecture
• @anakin87 for inspiring the idea
• The open-source AI community