Downloads · 30 days
20
10% of all-time downloads
LFM2.5-1.2B-Thinking-EXL3-4bpw
0xzknw/LFM2.5-1.2B-Thinking-EXL3-4bpw
LFM2.5-1.2B-Thinking-EXL3-4bpw is a text generation model from 0xzknw. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as other.
Local conversion of LiquidAI/LFM2.5-1.2B-Thinking, source revision f313478934a7612d22991f752959d7a1a8756fec. The original license is included in LICENSE.
All-time downloads
201
Public
Parameters
394M
788 MB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors788 MB · 99%
Parameter types
How the weights are stored.
U16259M · 66%
Try a prompt
Base models
Model card
From the Hugging Face model README
LFM2.5-1.2B-Thinking — EXL3 4 bpw
Local conversion of LiquidAI/LFM2.5-1.2B-Thinking,
source revision f313478934a7612d22991f752959d7a1a8756fec.
The original license is included in LICENSE.
- 92/92 linear projections converted to EXL3 K=4, MCG codebook.
- Tied embeddings, norms and convolution kernels retained in FP16.
- 4 bpw describes the quantized linear weights, not the complete checkpoint.
- Fresh calibration: 2,048 activation rows, 1,024-token sequences.
- Metal LDLQ, 128-row buffers, 16-row feedback, sigma 0.025, no QAT.
- Original tokenizer, chat template and generation configuration preserved.
Held-out WikiText-2 test, 2,048 scored tokens in 256-token windows: BF16 perplexity 265.253; EXL3 perplexity 271.717 (+2.44%). This small raw-text evaluation is not a general reasoning-quality benchmark.
Run with MLXL3
Install MLXL3, then download and register this checkpoint:
hf download 0xzknw/LFM2.5-1.2B-Thinking-EXL3-4bpw --local-dir ./LFM2.5-1.2B-Thinking-EXL3-4bpw
mlxl3 register lfm2.5-1.2b-thinking-4bpw ./LFM2.5-1.2B-Thinking-EXL3-4bpw
mlxl3 run lfm2.5-1.2b-thinking-4bpw
- Task
- Text Generation
- Library
- mlx
- Type
- lfm2
- License
- other
- Created
- Sep 8, 2026
- Updated
- Sep 8, 2026