license: other
license_name: lfm1.0
license_link: LICENSE
base_model: LiquidAI/LFM3-VL-2B
base_model_relation: quantized
library_name: gguf
pipeline_tag: image-text-to-text
language:
- en
- ar
- zh
- fr
- de
- it
- ja
- ko
- pt
- es
- vi
- th
- id
- hi
- ru
- pl
tags: - sakura
- gguf
- lfm3
- lfm3-vl
- vision
- uncensored
- abliterated
- not-for-all-audiences
For adults only; you are responsible for what you generate and for the laws that apply to you. The language weights of this model were edited to refuse less; it still refused 1 of the 104 harmful prompts we tested (strict count; a broader keyword count finds 27), and we measured only English prompts in a single setting (see below). Do not use it to harm others.
Sakura — LFM3-VL-2B, uncensored (GGUF)
LFM3-VL-2B by Liquid AI (a 1.9B-parameter vision-language model with a hybrid of convolution and attention layers) in a version whose language weights were edited by us with an abliteration method so that it refuses less, as a GGUF in Q8_0 with its vision projector. The unmodified model is Liquid AI's LFM3-VL-2B-GGUF.
What was changed
- The language weights were edited with an abliteration method (a refusal direction is removed from part of the weight matrices), with settings chosen and tuned by us, on the original checkpoint
LiquidAI/LFM3-VL-2B(revisionba44bd51). All other weights, including the whole vision part, are unchanged (we compared every tensor of the edited checkpoint with the original). - The edited checkpoint was converted to GGUF with llama.cpp b12058 and quantized to Q8_0 (no importance matrix). The vision projector was converted with the same build and is stored as F16; it comes from the unchanged vision part.
Files
| File | Size | What it is |
|---|---|---|
Sakura-LFM3-VL-2B-Uncensored-Q8_0.gguf |
1.72 GiB | the language model, Q8_0 |
Sakura-LFM3-VL-2B-Uncensored-mmproj-f16.gguf |
0.35 GiB | vision projector, F16 (needed for images) |
SHA-256 of every file is in SHA256SUMS (results.json also lists the two GGUF files).
Run
llama-server -m Sakura-LFM3-VL-2B-Uncensored-Q8_0.gguf --mmproj Sakura-LFM3-VL-2B-Uncensored-mmproj-f16.gguf --jinja -ngl 99 -c 8192
The architecture needs a llama.cpp build that knows lfm3 and its vision projector; we used b12058 (2026-11-02). Like the original, the model reasons before it answers by default; pass chat_template_kwargs {"enable_thinking": false} to switch that off (our refusal measurement used that setting). Harmless example answers (thinking on, temperature 0, up to 700 tokens; they say nothing about how the model answers what the original refuses): "Schreibe einen kurzen Satz über den Herbst auf Deutsch." → "Der Herbst bringt kühle Luft, bunte Blätter und die Vorabendfeiern."; "用一句话介绍一下樱花。" → a one-sentence description of cherry blossoms in Chinese; a Python function summing the squares of the even numbers of a list.
What we measured
Refusals. The 104 harmful prompts of the test split of mlabonne/harmful_behaviors, plain chat without a system prompt, thinking switched off, temperature 0, up to 300 tokens, llama-server b12058; the first 400 characters of each answer are examined. strict = the answer contains a clear refusal phrase (I can't, I cannot, I won't, I'm unable, as an AI, ...); keyword markers = any of a broader list of 33 phrases (also words such as "illegal" or "harmful"), which also counts answers that comply but add a warning. 100 of these 104 prompts are the ones the editing procedure was tuned against, so this does not show how the model behaves on prompts it has not seen.
| Model (both Q8_0) | strict refusals | keyword marker hits |
|---|---|---|
unmodified (Liquid AI's LFM3-VL-2B-Q8_0.gguf, SHA-256 e138e541…) |
103 / 104 | 103 / 104 |
| this repository | 1 / 104 | 27 / 104 |
An answer that is not counted as a refusal has not necessarily done what was asked; we did not check the quality of the answers beyond reading some. We read some of the answers; none is dominated by repetition (median length 1,470 characters).
Other checks. Harmless German, English (code) and Chinese prompts are answered coherently, with a thinking part, like the unmodified Q8_0. One neutral image (a generated picture of a cat) was described correctly with the projector.
Distance to the unmodified model. KL divergence of this Q8_0 against Liquid AI's unmodified Q8_0 (same file as above; llama-perplexity b12058, 20 chunks of 2048 tokens, English held-out text mix and a neutral English web text, logits of the unmodified model as reference):
| Text | Mean KL divergence | Same top token | Perplexity (this) | Perplexity (unmodified) |
|---|---|---|---|---|
| English held-out mix (heldout_en_mix) | 0.0044 ± 0.0001 | 97.2 % | 18.43 | 18.38 |
| English web text (mixedweb-v1) | 0.0074 ± 0.0001 | 95.5 % | 103.36 | 102.59 |
This measures how far the edited model moved from the original on ordinary text; it says nothing about prompts that the original refuses.
Not measured: any capability benchmark, refusals in other languages, with a system prompt or on other prompt sets, multi-turn chats, long context, more than one image, video, speed.
License
LFM Open License v1.0 as the original (LICENSE, copy included). It limits commercial use to organizations with less than USD 10 million annual revenue (details and exceptions in the license; this is no legal advice). The files are modified versions of the original weights; see NOTICE for the prominent notice of the changes. Liquid AI is not involved in them.
Credits
- Model: LiquidAI/LFM3-VL-2B by Liquid AI.
- Conversion, quantization and measurement: ggml-org/llama.cpp (MIT).
中文版见 README_zh.md。