Runs on — · Hugging Face ↗
Not independently scored by Artificial Analysis — its base model LiquidAI/LFM2.5-24B-A2B-Base is the closest proxy
Released by Liquid AI under the LFM Open License v1.0: commercial use only for companies under 10 million dollars of annual revenue. Text only, a mixture of experts with 2.45B active: 40 layers, 30 of them short-convolution blocks and 10 grouped-query attention, 64 experts with 4 active. Same architecture as LFM2-24B-A2B, with 131K context instead of 32K and a 128K vocabulary. It always thinks before it answers.
Real GGUF file sizes = weight VRAM.
| Quantization | Size |
|---|---|
| Q4_0 | 13.6 GB |
| Q4_K_M | 14.5 GB |
| Q5_K_M | 17.0 GB |
| Q6_K | 19.7 GB |
| Q8_0 | 25.5 GB |
| BF16 | 48.0 GB |
| F16 | 48.0 GB |
The Q4_K_M is 14.5 GB. The repo also has a QAD Q4_0 (13.6 GB), distilled with the quantization in place to recover accuracy. This site does not compute the KV cache of its convolution layers yet, so it does not say which cards it fits. Liquid AI recommends temperature 0.1, top_k 50 and repeat penalty 1.05, and the GGUF carries them as defaults. Liquid AI does not recommend it for knowledge-heavy tasks without retrieval.