LFM2.5-24B-A2B-Uncensored-NVFP4

NVFP4 (W4A16) quantization of pottokao/LFM2.5-24B-A2B-Uncensored (LiquidAI LFM2.5-24B-A2B decensored with Heretic ARA, searched and tested with thinking enabled; see the main repo for results).

• 14.95 GB (BF16: 48 GB).
• Only the routed MoE experts are NVFP4 (W4A16, group size 16, RTN, no calibration data, NVIDIA ModelOpt 0.45). Everything else stays BF16: attention, short-conv blocks, the two dense layers, router gates, embeddings and lm_head — including the two edited tensors (model.layers.22/26.self_attn.out_proj.weight, bit-identical to the BF16 release).
• w1/w3 of each expert share one per-tensor scale, because vLLM fuses them into one matrix. Expert tensors use the native w1/w2/w3 names.

Serving

vllm serve pottokao/LFM2.5-24B-A2B-Uncensored-NVFP4

The format follows the same layout as other ModelOpt W4A16 LFM2.5 checkpoints that vLLM detects as modelopt_mixed. Tested on vLLM (nightly fe4df3df, Marlin weight-only FP4 on GB10): loads, thinks, answers in Traditional Chinese and English, and calls tools correctly.

Liquid AI's recommended sampling: temperature 0.1, top_k 50, repetition_penalty 1.05.

⚠️ Notice

Refusal behavior has been substantially reduced. Intended for research, red-teaming and personal use; you are responsible for your use and for complying with laws and the base model license.

License & credits

Derivative of LiquidAI/LFM2.5-24B-A2B under the LFM Open License v1.0 (see LICENSE). Decensoring with Heretic (p-e-w, AGPL-3.0) / ARA. Quantization with NVIDIA ModelOpt.