🤖 LFM2.5-VL-1.6B-heretic — Importance Matrix GGUF
This repository hosts importance-matrix (imatrix) optimized GGUF weights, available in multiple quantization formats, and the associated vision projection matrix for
LFM2.5-VL-1.6B-heretic, quantized from the source floating-point tensors provided by
coder3101/LFM2.5-VL-1.6B-heretic.
🔄 Sister Repository: Check out the
Standard GGUF Sister Repository for uncalibrated and full 8-bit precision options.
🎯 Matrix-Weighted Calibration (Imatrix)
An Importance Matrix (imatrix) calculation tracks activations across network layers using a calibration sequence, then weights the quantization process to preserve the parameters that matter most for output quality — improving fidelity at low bit depths.
➡️ Calibration dataset: Bartowski's calibration_datav5.txt.
[!NOTE]
IQ4_NL is included because the matrix enables a non-linear 4-bit format that outperforms standard linear 4-bit quantization.Q8_0 is absent because 8-bit quantization already introduces near-zero degradation, making calibration unnecessary — see the standard sister repository for that variant.
ℹ️ Model Profile & Core Features
LFM2.5-VL-1.6B is the larger vision-language model in Liquid AI's LFM2.5-VL family, pairing the LFM2 hybrid language backbone (gated short convolutions interleaved with GQA) with a SigLIP2 NaFlex vision encoder for single- and multi-image understanding. It is designed to run comfortably on a single commodity GPU while supporting a 32,768-token context window, and ships with day-one support for llama.cpp, vLLM, and MLX.
The
heretic suffix denotes post-processing via the
Heretic v1.3.0 method performed by
coder3101, which suppresses refusal behavior while preserving the model's vision-language capabilities.
📋 Technical Specifications
| Property | Value |
|---|
| Base Architecture | LFM2 hybrid (gated conv + GQA) + SigLIP2 NaFlex vision encoder |
| Developed by | Liquid AI |
| Total Parameters | 1.6B (LM + vision encoder) |
| Vision Encoder | SigLIP2 NaFlex, shape-optimized 400M |
| Primary Use | Single/multi-image understanding, visual Q&A |
| Context Window | 32,768 tokens |
| Vision Projector | Integrated (see repository files below) |
| Languages | English, Japanese, Korean, French, Spanish, German, Arabic, Chinese |
| Abliteration Tool | Heretic v1.3.0 |
| Prompt Format | ChatML |
🛠️ Heretic Overrides (ARA)
| Property | Value |
|---|
| direction_index | 10.24 |
| attn.o_proj.max_weight | 1.22 |
| attn.o_proj.max_weight_position | 10.01 |
| attn.o_proj.min_weight | 1.21 |
| attn.o_proj.min_weight_distance | 7.36 |
| mlp.down_proj.max_weight | 1.11 |
| mlp.down_proj.max_weight_position | 12.44 |
| mlp.down_proj.min_weight | 0.21 |
| mlp.down_proj.min_weight_distance | 5.13 |
📊 Refusal Bypass Metrics
[!NOTE]
The metrics below are self-reported by the original model author (
coder3101) and have not been independently reproduced.
🧮 Numerical & Tensor Formats
| Property | Value |
|---|
| Text Tensor Types | IQ4_NL, Q4_K_M, Q5_K_M (all with imatrix calibration) |
| Importance Matrix | Bartowski's calibration_datav5.txt |
| Vision Tensors | Q8_0, BF16 |
📦 Available Model Files
Main model weights
| Filename | Quantization | llama.cpp Build | Size | Download |
|---|
LFM2.5-VL-1.6B-heretic-IQ4_NL-imatrix.gguf | IQ4_NL | b9860 | 664 MB | 📥 Download |
LFM2.5-VL-1.6B-heretic-Q4_K_M-imatrix.gguf | Q4_K_M | b9860 | 697 MB | 📥 Download |
LFM2.5-VL-1.6B-heretic-Q5_K_M-imatrix.gguf | Q5_K_M | b9860 | 804 MB | 📥 Download |
mmproj — vision projector files
| Filename | Quantization | Size | Download |
|---|
mmproj-LFM2.5-VL-1.6B-heretic-Q8_0.gguf | Q8_0 | 556 MB | 📥 Download |
mmproj-LFM2.5-VL-1.6B-heretic-BF16.gguf | BF16 | 816 MB | 📥 Download |
🎛️ Component Pairing Guide
Download exactly one main weights file:
IQ4_NL: Non-linear 4-bit format, best choice for constrained memory when imatrix calibration is present.Q4_K_M: Balanced 4-bit format suitable for most everyday use.Q5_K_M: Higher-fidelity mid-range format recommended as a general default.
mmproj files (optional): multimodal vision projectors. Pass one via the --mmproj flag in llama.cpp to enable image input.
BF16 (Recommended): Highest possible image processing accuracy. While older projectors were small, modern vision towers can hover around 1GB. If you are tight on VRAM, it is completely viable to run this on system RAM (CPU) with a minimal performance penalty, saving your precious GPU space for the main model layers.Q8_0: Cuts the projector file size and memory footprint in half (~500MB for larger 1GB files). Use this if you prefer to keep the vision tower hosted entirely on your GPU but need to claw back some VRAM to avoid Out-Of-Memory (OOM) crashes.
⚡ Deployment & Execution Commands
[!IMPORTANT]
The vision projector (--mmproj) must be supplied at runtime whenever image inputs are used. Omitting it disables multimodal capability entirely.
[!NOTE]
LiquidAI recommends the following sampling configuration for best results:
- Text:
temperature=0.1, min_p=0.15, repetition_penalty=1.05.
- Vision:
min_image_tokens=64, max_image_tokens=256, do_image_splitting=True.
[!TIP]
Swap the -m filename below for either quantized file depending on your size/quality trade-off preference.
llama.cpp CLI (with image)
1./llama-cli \
2 -m LFM2.5-VL-1.6B-heretic-IQ4_NL-imatrix.gguf \
3 --mmproj mmproj-LFM2.5-VL-1.6B-heretic-Q8_0.gguf \
4 -c 8192 \
5 -ngl 99 \
6 --image "path/to/image.jpg" \
7 -p "<|im_start|>user\nDescribe what you see in this image.<|im_end|>\n<|im_start|>assistant\n"
OpenAI-Compatible API Server
1./llama-server \
2 --host 0.0.0.0 \
3 --port 8080 \
4 -m LFM2.5-VL-1.6B-heretic-IQ4_NL-imatrix.gguf \
5 --mmproj mmproj-LFM2.5-VL-1.6B-heretic-Q8_0.gguf \
6 -c 16384 \
7 -ngl 99 \
8 --flash-attn
💬 Chat Templates & Prompt Design (ChatML)
1<|im_start|>system
2You are a helpful multimodal assistant.<|im_end|>
3<|im_start|>user
4Your question or image payload here.<|im_end|>
5<|im_start|>assistant
⚠️ Safety & Operational Notes
- This model is abliterated and will generate content that standard aligned models refuse. Use responsibly and in compliance with applicable laws.
- Fits comfortably on a single GPU with at least 8 GB VRAM at quantized precision.
- Context window is limited to 32,768 tokens — shorter than the text-only LFM2.5 models.
- Vision tensors are kept at BF16 or Q8_0 depending on the mmproj variant chosen, to preserve visual feature quality.
- Imatrix calibration improves perplexity recovery compared to non-imatrix quantization, particularly on low-frequency tokens.
- IQ4_NL produces a smaller file than Q4_K_M and tends to run faster on CPU and ARM devices; imatrix calibration narrows the quality gap between the two formats considerably.
