konic-labs/LFM2.5-VL-3B-ffn8192
3B • Updated • 12
Optimized by Konic Labs: LFM2.5 3B Vision model's three near-lossless versions down to 2.57 GB INT4 (−69%)
Note Pruned 10752->8192 + distillation recovery, BF16 5.30 GB (-15%)
Note Pruned 10752->7168 + distillation recovery, BF16 4.93 GB (-21%)
Note Native INT4/INT8 compressed-tensors, 2.57 GB (-69%), vLLM-ready
Note Bonus: FFN 7168 + INT4/INT8, 2.50 GB (-60%). Caveat: tool-call closing tag may be missing.
Note Bonus: deepest FFN prune 6144, BF16 4.57 GB (-27%). Caveat: tool-call closing tag may be missing.
Note Smallest: FFN 6144 + INT4/INT8, 2.40 GB (-62%). Caveat: tool-call closing tag may be missing.