language:
- en
- zh
base_model: LiquidAI/LFM3-VL-2B
pipeline_tag: image-text-to-text
tags: - lfm3_vl
- abliterix
- edge
license: other
license_name: lfm1.0
license_link: LICENSE
LFM3-VL-2B-Uncensored-MLX
A vision-language model based on LiquidAI/LFM3-VL-2B, modified with Abliterix to reduce refusal behavior, for local and edge-device deployment.
The model supports text and image inputs, retaining the vision encoder and multimodal projector while modifying the language backbone's attention weights with layer-specific refusal directions.
Model details
| Property | Description |
|---|---|
| Base model | LiquidAI/LFM3-VL-2B |
| Task | Image-text-to-text |
| Context length | Up to 128,000 tokens, subject to runtime settings and available memory |
| Modification | Abliterix; language attention Q/K/V/O projections |
| License | LFM Open License v1.0; see LICENSE |
Abliterix uses layer-specific mean-difference directions from 400 benign and 400 target training prompts, direct orthogonal projection and pre-edit weight normalization.
MLX weights
| Location | Precision |
|---|---|
| Repository root | 4-bit |
6bit/ |
6-bit |
8bit/ |
8-bit |
BF16/ |
Unquantized BF16 |
Integer variants use affine group64 quantization of the language model; the vision encoder and multimodal projector remain BF16. These are post-training conversions, not the upstream quantization-aware-distilled checkpoints.
Usage
For Apple Silicon, use MLX-VLM with LFM3-VL support, such as version 0.7.11 or later. Load the complete weights, configuration, tokenizer and image processor from the chosen precision directory. Strict CPU architecture reload and tensor-integrity checks were completed; Apple-device inference was not tested.
Other formats: BF16 and compressed-tensors · GGUF.
Suggested sampling settings: temperature=0.1, top_p=1.0, top_k=50, min_p=0.0, repetition_penalty=1.1.
Evaluation
These metrics describe the BF16 source model; quantized variants have not been separately benchmarked for them.
| Metric | Result |
|---|---|
| Full-vocabulary next-token KL relative to the base model | 0.0426511 |
| StrongREJECT strict refusal rate | 4% |
| StrongREJECT warning while still answering | 38% |
| SimpleSafetyTests strict refusal rate | 8% |
| SimpleSafetyTests warning while still answering | 56% |
KL is mean KL(P_base || P_modified) at one next-token position for each of 100 fixed benign prompts, over all 128,000 vocabulary entries; it is not whole-sequence divergence.
StrongREJECT uses a fixed stratified 100-item subset; SimpleSafetyTests uses all 100 cases. The judge is OllamaCloud deepseek-v4.1-flash:1017. A response that warns but substantively answers is counted as answering, not refusal. Both evaluations had zero generation errors, judge errors, unresolved labels and truncated responses. Free-form generation had no explicit output-token cap, within the context limit.
Abliterix and quantization may affect response quality. Evaluate the exact format, runtime and task before deployment.
License
Derived from LiquidAI/LFM3-VL-2B under its original LFM Open License v1.0.