---
license: other
language:
- en
- zh
library_name: llama.cpp
pipeline_tag: text-generation
tags:
- gguf
- lfm2
- lfm2.5
- llama.cpp
- atomic-chat
base_model:
- LiquidAI/LFM2.5-2.6B
---

# LFM2.5 2.6B Uncensored — GGUF

Unofficial GGUF conversions of [SC117/LFM2.5-2.6B-Uncensored](https://huggingface.co/SC117/LFM2.5-2.6B-Uncensored), prepared from the original BF16 weights for local inference with `llama.cpp`, Atomic Chat, LM Studio, and other GGUF-compatible applications.

This repository is not affiliated with Liquid AI. Please review the upstream [license](https://huggingface.co/SC117/LFM2.5-2.6B-Uncensored/blob/main/LICENSE) before redistribution or commercial use.

## Available files

| Quantization | File size | Recommendation |
|---|---:|---|
| Q4_K_M | 1.67 GB | Smallest; fastest and lowest memory use |
| Q5_K_M | 1.94 GB | Good speed/quality balance |
| Q6_K | 2.22 GB | Recommended default for quality-sensitive prompt generation |
| Q8_0 | 2.87 GB | Closest to BF16 quality; use when quality is the priority |

Sizes are the local GGUF file sizes in decimal GB. All variants retain the model's tokenizer, chat template, and metadata.

### File SHA-256

| File | SHA-256 |
|---|---|
| `LFM2.5-2.6B-Uncensored-Q6_K.gguf` | `7CDDB332D4284B2113AD89E7D1C6FDC968EDBAD2704CB05A76D439BB98E06EAF` |
| `LFM2.5-2.6B-Uncensored-Q8_0.gguf` | `61A09E86309A80522E7E65D5FC0A28798148CA9F388630C390210D3676E03434` |

## Recommended choice for an M2 MacBook with 16 GB RAM

- **Q6_K** is the recommended everyday choice for short English/Chinese image and video prompts.
- **Q8_0** is also practical and should preserve a little more quality, at the cost of a larger model and somewhat higher memory bandwidth use.
- Keep context moderate (for example, 4k–8k tokens) for the fastest response.

The model weights fit comfortably in 16 GB unified memory; the operating system, application, context, and KV cache still need headroom.

## llama.cpp

```bash
llama-server \
  -m LFM2.5-2.6B-Uncensored-Q6_K.gguf \
  -c 8192 -ngl 99 -fa on --jinja
```

Use the Q8 file in the `-m` argument when maximum quality is preferred. The exact flags available depend on the llama.cpp build and hardware backend.

## Atomic Chat

Load one of the GGUF files with the **Llama.cpp** engine. For structured prompt generation, Atomic Chat's Llama.cpp settings expose **Grammar File** and **JSON Schema File** output constraints. A simple flat JSON Schema is usually more reliable than a large schema with nested references.

Atomic Chat's local OpenAI-compatible server is normally available at `http://127.0.0.1:1337/v1`. With a llama.cpp backend, a request can use `response_format` with `type: "json_schema"` to constrain the generated prompt object. Do not combine a custom grammar/schema with native tool-call grammar in the same request.

## Example structured output schema

```json
{
  "type": "object",
  "properties": {
    "prompt_en": { "type": "string" },
    "prompt_zh": { "type": "string" },
    "negative_prompt": { "type": "string" },
    "parameters": { "type": "string" }
  },
  "required": ["prompt_en", "prompt_zh", "negative_prompt", "parameters"],
  "additionalProperties": false
}
```

The schema constrains the output shape; describe the intended fields in the prompt as well. Validate the returned JSON in the calling application.

## Quantization provenance

The files were quantized directly from the BF16 GGUF conversion of the upstream model with `llama-quantize` (llama.cpp build `10278`, commit `d52ec04a6`):

```bash
llama-quantize LFM2.5-2.6B-Uncensored-BF16.gguf LFM2.5-2.6B-Uncensored-Q4_K_M.gguf Q4_K_M 8
llama-quantize LFM2.5-2.6B-Uncensored-BF16.gguf LFM2.5-2.6B-Uncensored-Q5_K_M.gguf Q5_K_M 8
llama-quantize LFM2.5-2.6B-Uncensored-BF16.gguf LFM2.5-2.6B-Uncensored-Q6_K.gguf Q6_K 8
llama-quantize LFM2.5-2.6B-Uncensored-BF16.gguf LFM2.5-2.6B-Uncensored-Q8_0.gguf Q8_0 8
```

Original model: [SC117/LFM2.5-2.6B-Uncensored](https://huggingface.co/SC117/LFM2.5-2.6B-Uncensored)