---
title: LFM3-1.7B-GGUF
canonical_url: "https://www.modelscope.ai/models/atomic.chat/LFM3-1.7B-GGUF"
md_url: "https://www.modelscope.ai/models/atomic.chat/LFM3-1.7B-GGUF.md"
repository: atomic.chat/LFM3-1.7B-GGUF
last_updated: 2026-10-28
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
base_model:
  - LiquidAI/LFM3-1.7B
base_model_relation: quantized
library_name:
  - gguf
  - pytorch
frameworks:
  - pytorch
language:
  - en
  - zh
  - de
  - fr
  - es
  - ja
  - ar
  - ko
downloads: 8
stars: 0
tags:
  - gguf
  - quantized
  - lfm3
---

# LFM3-1.7B-GGUF

> LFM3-1.7B-GGUF - An open-source model by atomic.chat on ModelScope. LFM3-1.7B — GGUF (by AtomicChat)

atomic.chat/LFM3-1.7B-GGUF is a text-generation model on ModelScope. licensed under other. derived from LiquidAI/LFM3-1.7B.

- **Repository**: atomic.chat/LFM3-1.7B-GGUF
- **License**: other
- **Tasks**: text-generation
- **Base model**: LiquidAI/LFM3-1.7B
- **Tags**: gguf, quantized, lfm3
- **Downloads**: 8
- **Stars**: 0
- **Last updated**: 2026-10-28

Source: https://www.modelscope.ai/models/atomic.chat/LFM3-1.7B-GGUF

---

# LFM3-1.7B — GGUF (by AtomicChat)

Atomic-Dynamic (AD) GGUF quants of [LiquidAI/LFM3-1.7B](https://huggingface.co/LiquidAI/LFM3-1.7B), a dense short-convolution + gated-attention hybrid.

Each AD rung sets per-tensor types by role (FFN pressed hardest by depth band, `attn_gate` and `attn_k/v` kept high, the short-conv mixers tracked with attention, embeddings and head high) rather than a single uniform preset.

## Rungs

`top-1` is how often the rung picks the same next token as the original BF16 (30-language neutral set); mean KLD is the average divergence from BF16 (lower is better). LFM3-1.7B is a small hybrid model and is fairly quantization-sensitive.

| Rung | GB on disk | bpw | top-1 | mean KLD |
|---|---|---|---|---|
| Q8_0 | 1.85 | 8.51 | 96.3% | 0.004 |
| AD-Q6_K | 1.59 | 7.31 | 94.5% | 0.010 |
| AD-Q5_K_M | 1.31 | 6.03 | 90.3% | 0.034 |
| AD-Q4_K_M | 1.13 | 5.20 | 84.6% | 0.087 |
| AD-IQ4_XS | 1.09 | 5.02 | 83.2% | 0.105 |
| AD-IQ3_S | 0.92 | 4.23 | 77.0% | 0.203 |

**AD-Q4_K_M** is the recommended default. **AD-IQ3_S** is a research-grade rung (small, lower fidelity). The AD rungs replace llama.cpp's stock presets, which have no imatrix, at every size. Full metrics go to the `-metrics` repo after audit.

## Run (llama.cpp)

Needs a build with LFM3 support (merged in llama.cpp #31252; these files were produced at llama.cpp `19cce69`).

```bash
llama-cli   -hf AtomicChat/LFM3-1.7B-GGUF:AD-Q4_K_M -p "Hello"
llama-server -hf AtomicChat/LFM3-1.7B-GGUF:AD-Q4_K_M -c 8192 --jinja
```

## Base model & license

Base model: [LiquidAI/LFM3-1.7B](https://huggingface.co/LiquidAI/LFM3-1.7B). License: LFM Open License v1.0 (see `LICENSE`).

Quants by AtomicChat.