ABLITERIX TRIAL 65 VISION GGUF + IMATRIX LFM Open 1.0

LFM2.5-VL-3B-Uncensored-GGUF

English | 📖 中文文档

Uncensored vision-language edge model · abliterix Trial 65 merged into official LFM2.5-VL-3B · imatrix-calibrated GGUFs + mmproj

🌊 About this release

LFM2.5-VL-3B is Liquid AI's official 3.1B vision-language edge model: the LFM2.5-2.6B hybrid language backbone (30 layers: 22 double-gated short-convolution blocks + 8 GQA, 128K context, 128K vocab) paired with a SigLIP2 NaFlex vision encoder (256×256, patch 16) and a language-aligned projector.

This release merges the abliterix Trial 65 LoRA (from our LFM2.5-2.6B-Uncensored, rank-1, alpha = r = 1) into the language-model layers of the official VL checkpoint (q/k/v/out_proj, feed_forward.w2, conv.out_proj — 84 tensors across 30 layers). The vision encoder and projector are untouched, so multimodal capability is fully preserved.

Build steps:

1. LoRA merge into the VL checkpoint: W += (B @ A) * (alpha / r), applied to model.language_model.layers.*.
2. BF16 GGUF conversion of the merged model (llama.cpp lfm2 architecture) + mmproj export of the vision tower via convert_hf_to_gguf.py --mmproj --outtype bf16.
3. imatrix calibration — reused the 401-chunk (≈1.6M tokens) importance matrix from the 2.6B release (same architecture, weights closely related).
4. Quantization with llama-quantize --imatrix into five tiers.

License: LFM Open License v1.0 (same as the base model).

⚠️ Uncensored notice

After merging the Trial 65 steering, this model shows a much lower refusal rate on both text and vision-grounded prompts, and can differ substantially from official LFM2.5-VL-3B. Evaluate compliance and safety for your use case; control access and audit as needed.

📦 Files

Main model (language backbone, lfm2 architecture, 128K context) — pick one tier and pair it with the mmproj:

Vision tower (required, pick one):

All main-model tiers are lfm2 architecture, 128K context, single-file GGUFs, imatrix-calibrated.

💡 Why imatrix?

The importance matrix (401 chunks / ≈1.6M tokens of mixed conversation, math, and code data, computed on the 2.6B-Uncensored BF16 GGUF) tells the quantizer which weights are sensitive. Since the VL language backbone shares the exact same architecture and near-identical weights, the matrix transfers cleanly (166/166 quantized tensors matched; only token_embd falls back to plain q6_K). K-quants and especially the IQ tiers use it to keep more bits on attention/embedding paths.

🚀 Usage (llama.cpp)

Pair any main-model tier with the mmproj (recent llama.cpp, e.g. b10299+, required for lfm2 VL support):

llama-server -m LFM2.5-VL-3B-Uncensored-Q4_K_M.gguf \
  --mmproj mmproj-LFM2.5-VL-3B-BF16.gguf \
  --ctx-size 8192 --flash-attn on --host 0.0.0.0 --port 8080

CLI with an image:

llama-llava-cli -m LFM2.5-VL-3B-Uncensored-Q4_K_M.gguf \
  --mmproj mmproj-LFM2.5-VL-3B-BF16.gguf \
  -i image.jpg -p "Describe this image." -ngl 99

Transformers / vLLM / SGLang users: use the merged BF16 safetensors in the parent repo (to be published).

🧠 Behavior notes

• No thinking mode. Unlike the 2.6B text model, official LFM2.5-VL-3B is trained to answer directly for low-latency edge tasks; the <think> tags are not generated. Keep max_new_tokens modest.
• Vision strengths (per official model card): near-real-time object detection, OCR with layout annotation, document/chart understanding, on-device translation.
• Chat template: ChatML-like with <image> placeholder, same as official. Recommended sampling: temperature 0.2, top_k 50, repetition_penalty 1.0.

📊 Benchmarks (official model card)

Values below are taken from the official LiquidAI model card for LFM2.5-VL-3B (not re-measured on this release):

Community derivative (behavior edit + quantized GGUF release). Not an official Liquid AI release. Use at your own risk; follow local law and the LFM Open License v1.0.