LFM2.5-8B-A1B-REFFT
Hexagon HTP serving pack for
LiquidAI/LFM2.5-8B-A1B, exported for refft-hexagon.
Artifact
LFM2.5-8B-A1B-w4.refft/ # directory serving pack (hexagon.serving.v1)
embed_f16.bin
world1/ # 1×HTP weight pages
world2/ # 2×HTP weight pages (rank0 + rank1)
programs/
metadata.json
serving_manifest.json
tokenizer*
- Quant: mixed W4 (+ FP16 embed)
- Device worlds:
1and2(both included; ~8G each + shared embed) - Load with product CLI:
refft-hexagon cli --model /path/to/LFM2.5-8B-A1B-w4.refft \
--backend hexagon --world_size 1 --prompt "Hello"
Dual HTP:
refft-hexagon cli --model /path/to/LFM2.5-8B-A1B-w4.refft \
--backend hexagon --world_size 2 --prompt "Hello"
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support