license: other
license_name: lfm-open-license-v1.0
license_link: LICENSE
base_model: OpenIntelligenceNet/LFM3-Omni-2B-Uncensored
base_model_relation: quantized
library_name: gguf
pipeline_tag: any-to-any
tags:
- gguf
- lfm3
- lfm3-omni
- multimodal
- vision
- audio
LFM3-Omni-2B-Uncensored — GGUF
GGUF exports of OpenIntelligenceNet/LFM3-Omni-2B-Uncensored,
which is based on LiquidAI/LFM3-Omni-2B.
These files were converted from the author's checkpoint at revision23b95a548729b23363732625d2f40f05f5c9a55f.
No additional training, merging or abliteration was performed for this export.
Files
Download one language-model GGUF and the matching Omni projector. Keep them
in the same folder.
| File | Precision | Purpose |
|---|---|---|
LFM3-Omni-2B-Uncensored-BF16.gguf |
BF16 | Unquantized language-model export |
LFM3-Omni-2B-Uncensored-Q8_0.gguf |
Q8_0 | Smaller, 8-bit language model |
LFM3-Omni-2B-Uncensored-Q4_K_M.gguf |
Q4_K_M | Smaller, mixed 4-bit language model; quantization can affect quality |
mmproj-LFM3-Omni-2B-Uncensored-BF16.gguf |
BF16 | Matching vision/audio projector, depth decoder and codec components |
The same projector is used with all three language-model variants. It was
converted from this same checkpoint, including its audio_tokenizer weights;
it is not borrowed from a different fine-tune. SHA256SUMS records file checksums.
Running
Use llama.cpp b12063 or newer with LFM3 support. An older version which only
supports LFM2 cannot load these files.
llama-server \
-m LFM3-Omni-2B-Uncensored-BF16.gguf \
--mmproj mmproj-LFM3-Omni-2B-Uncensored-BF16.gguf \
--reasoning off \
--ctx-size 8192
Substitute the Q8_0 or Q4_K_M filename to use a smaller model. The model's maximum
context is 32,768 tokens including input and output. LFM3-Omni is a non-reasoning
model. For text chat, use a system instruction such as Respond with text.
The original model documentation describes the available audio and interleaved
generation modes; their availability also depends on the client and runtime.
Conversion and validation
Conversion uses unmodified llama.cpp b12063
at commit 2cbca389f515df455412dafa07e2b1359cc72a6d. Q8_0 and Q4_K_M were
quantized directly from the BF16 export with the matching b12063 quantizer.
BF16 export is format conversion; the lower-bit quantizations are separate
files and do not replace it.
conversion.json records the source, converter and exported-file identities.
Validation results are recorded in VALIDATION.md. Test coverage should not be
interpreted as a guarantee that every upstream modality is supported by every
application. The original model card is preserved as UPSTREAM_README.md.
License and attribution
The model remains subject to the LFM Open License v1.0, included in LICENSE.
The format conversion does not replace the original model license with the
conversion tool's MIT license. See NOTICE for attribution.
Converted by brurpo for use with BRP Canvas 2
and compatible GGUF clients. Original model development is credited to Liquid AI
and the checkpoint modification to OpenIntelligenceNet.