LFM2.5-1.2B-Thinking-ToMoE-GGUF

LFM2.5-1.2B-Thinking-ToMoE-GGUF

GGUF quantizations of Nichonauta/LFM2.5-1.2B-Thinking-ToMoE — the ToMoE Mixture-of-Experts conversion of LiquidAI/LFM2.5-1.2B-Thinking.

Files

The Q4_K_M version runs under 1 GB of storage — matching the "on-device under 1 GB" goal of the original model.

Important note about the conversion

The ToMoE channel-MoE experts and the per-token dynamic V masks have no native representation in llama.cpp's LFM2 architecture. To produce a runnable GGUF, the pruned model was rebuilt as a dense-equivalent:

• The cut conv layers were restored to their full width using the original dense weights (the conv weights were never modified by ToMoE — only masked at runtime).
• The MLP union cut (~99% of the FFN was already preserved) was restored to the full FFN.
• The attention Q/K were already full-width in the pruned model.

As a result, the GGUF behaves like the dense base model rather than the pruned MoE (including the native <think> reasoning behavior of LFM2.5-Thinking). For the actual MoE behavior, use the safetensors version with trust_remote_code.

Usage (llama.cpp, CUDA)

llama-server.exe ^
  --model LFM2.5-1.2B-Thinking-ToMoE-Q4_K_M.gguf ^
  --gpu-layers all ^
  --ctx-size 32768 ^
  --temp 0.05 ^
  --top-k 50 ^
  --repeat-penalty 1.05 ^
  --alias LFM2.5-1.2B-Thinking-ToMoE

Or with llama-cli (single turn, shows the thinking block):

llama-cli.exe -m LFM2.5-1.2B-Thinking-ToMoE-Q4_K_M.gguf -ngl 99 -p "What is the capital of France?" -n 90

Measured on an RTX 3060 (CUDA 13.3 build, llama.cpp b10603): ~302 t/s generation for Q4_K_M, ~212 t/s for Q8_0, ~125 t/s for BF16. The official generation parameters of the base model (temperature 0.05, top_k 50, repetition_penalty 1.05) are recommended.

Base model

• MoE source: Nichonauta/LFM2.5-1.2B-Thinking-ToMoE
• Dense base: LiquidAI/LFM2.5-1.2B-Thinking (LFM Open License v1.0)

License

Derivative of LiquidAI/LFM2.5-1.2B-Thinking — released under the LFM Open License v1.0 (see LICENSE).