library_name: openvino
license: other
license_name: lfm1.0
license_link: LICENSE
base_model:
- LiquidAI/LFM2.5-24B-A2B
pipeline_tag: text-generation
tags: - openvino
- fp16
- bf16-source
- int8-moe
- intel
- arc
- lfm2.5
- moe
- uncensored
- abliterated
LFM2.5-24B-A2B-Uncensored BF16-source OpenVINO IR
GPU-loadable BF16-source/reference variant of the LoRA-fused, uncensored
LFM2.5-24B-A2B.
The source contains the rank-1 atomic.abliterate adapterr24-all-14-39-perlayer fused into layers 14–39 at scale 1.75. The adapter is
baked into the weights; no adapter is needed at runtime.
Precision note
OpenVINO IR has no portable BF16 weight-storage mode for this export path, so
the fused source is represented as FP16 IR. The convolution, attention,
normalization, router, embedding, and output-projection paths remain
uncompressed FP16. The 38 tiled MoE expert subgraphs use INT8 asymmetric
weight-only compression so the complete stateful graph fits the 32 GB Arc
iGPU memory limit. This is the high-precision BF16-source/reference variant;
the INT8 compression is isolated to the expert weight tensors.
Conversion
The source checkpoint was exported one decoder layer at a time because one-shot
export exceeds the machine's 30 GB RAM budget. The hybrid convolution and
attention state is preserved. MoE blocks were extracted from regular per-layer
OpenVINO exports, compressed only in their expert weights, and grafted with the
FP16 body layers into one stateful IR.
| Property | Value |
|---|---|
| Architecture | Lfm2MoeForCausalLM, 40 layers, 64 experts, top-4 routing |
| Weight format | FP16 body/reference; INT8 asymmetric MoE experts |
| Model size | 24 GB |
| Runtime | OpenVINO GenAI, GPU |
Usage
import openvino_genai as og
pipe = og.LLMPipeline("lfm2.5-24b-a2b-uncensored-bf16", "GPU")
config = og.GenerationConfig(max_new_tokens=256, temperature=0.7)
print(pipe.generate("Explain why the sky is blue.", config))
Benchmark
Single stream, Intel Core Ultra 7 258V Arc iGPU, OpenVINO GenAI 2026.4.0,
128-token generation with the same harness used for the INT4 and INT8 builds:
20.1 tok/s. The stateful path loads successfully on GPU; continuous
batching is unsupported for this hybrid convolution/state model.
Files
The repository follows the standard OpenVINO/OVMS layout: model IR, tokenizer,
detokenizer, HF config, generation config, graph.pbtxt, ovms_config.json,
license, and model documentation.
Attribution
- Base: LiquidAI/LFM2.5-24B-A2B
- Uncensoring: rank-1
atomic.abliterateLoRA,r24-all-14-39-perlayer - Conversion: OpenVINO IR, per-layer export, NNCF MoE compression, and graft
Redistribution follows the included LFM Open License v1.0.