LFM2.5-350M-ToMoE-GGUF
GGUF quantizations of Nichonauta/LFM2.5-350M-ToMoE — the ToMoE Mixture-of-Experts conversion of LiquidAI/LFM2.5-350M.
Files
| File | Quantization | Size | BPW |
|---|---|---|---|
LFM2.5-350M-ToMoE-BF16.gguf | BF16 | 678 MB | 16.0 |
LFM2.5-350M-ToMoE-Q8_0.gguf | Q8_0 | 362 MB | 8.50 |
LFM2.5-350M-ToMoE-Q4_K_M.gguf | Q4_K_M | 219 MB | 5.12 |
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. For the actual MoE behavior, use the safetensors version with trust_remote_code.
Usage (llama.cpp, CUDA)
llama-server.exe ^
--model LFM2.5-350M-ToMoE-Q4_K_M.gguf ^
--gpu-layers all ^
--ctx-size 32768 ^
--alias LFM2.5-350M-ToMoE
Or with llama-cli:
llama-cli.exe -m LFM2.5-350M-ToMoE-Q4_K_M.gguf -ngl 99 -p "The capital of France is" -n 30
Measured on an RTX 3060 (CUDA 13.3 build, llama.cpp b10603): ~470 t/s generation for Q4_K_M, ~370 t/s for Q8_0, ~270 t/s for BF16.
Base model
- MoE source: Nichonauta/LFM2.5-350M-ToMoE
- Dense base: LiquidAI/LFM2.5-350M (LFM Open License v1.0)
License
Derivative of LiquidAI/LFM2.5-350M — released under the LFM Open License v1.0 (see LICENSE).