---
license: other
license_name: lfm-open-license-v1.0
license_link: https://www.liquid.ai/lfm-open-license
base_model: LiquidAI/LFM2.5-8B-A1B
tags:
  - gguf
  - nobodywho
  - tool-calling
  - lfm2
  - moe
pipeline_tag: text-generation
library_name: gguf
---

# NobodyWho/LFM2.5-8B-A1B-GGUF

## Overview

GGUF quantization of LiquidAI's **LFM2.5-8B-A1B** model, prepared for
[NobodyWho](https://github.com/nobodywho-ooo/nobodywho): it works with NobodyWho
out of the box, with LiquidAI's recommended sampling metadata embedded in every quant.
LFM2.5-8B-A1B is a sparse Mixture-of-Experts model (8B total / ≈1B active per token) built on the
hybrid LFM2 architecture — the fastest model in its size class on both CPU and GPU.
**Note:** tool calling is unreliable on this model — see the Tool calling note below.

## Model Capabilities

- **Text generation** — instruction-following chat
- **Tool calling** — supported but **unreliable**: the model often answers in prose instead of
  invoking a tool. NobodyWho suite: 8/14 (F16, Q8_0), 4/14 (Q4_K_M)
- **Long context** — 128k tokens
- **Efficient MoE** — 8B total / ≈1B active per token

## NobodyWho preparation

The upstream GGUF (built from LiquidAI commit `feb5e04`) already renders tool calls correctly in
the model's native markup — `<|tool_call_start|>[get_weather(city="Paris")]<|tool_call_end|>` —
so nothing needs patching; NobodyWho just verifies it with the test suite. The
`-vendor-sampling` quants additionally embed LiquidAI's recommended sampling settings as
`general.sampling.*` metadata, which NobodyWho reads and applies by default (see
`core/src/sampler.rs`).

## Available Quantizations

| File | Approach | Tool-calling tests |
|------|----------|--------------------|
| `LFM2.5-8B-A1B-F16-vendor-sampling.gguf` | Vendor sampling injected | 8/14 |
| `LFM2.5-8B-A1B-Q8_0-vendor-sampling.gguf` | Vendor sampling injected | 8/14 |
| `LFM2.5-8B-A1B-Q4_K_M-vendor-sampling.gguf` | Vendor sampling injected | 4/14 |

> Tool-calling results from NobodyWho's suite (June 2026). Failures are the model **declining to
> emit a tool call** on complex parameter schemas (sets / tuples / nested lists / dicts), not a
> format error. Vendor sampling does **not** change the result (verified with and without).
> The `-vendor-sampling` suffix marks files that embed `general.sampling.*` metadata.

## Quick Start

Using the [NobodyWho](https://github.com/nobodywho-ooo/nobodywho) library:

```python
from nobodywho import Chat

chat = Chat("huggingface:NobodyWho/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0-vendor-sampling.gguf")
response = chat.ask("What is the capital of Denmark?").completed()
print(response)  # The capital of Denmark is Copenhagen.
```

### llama-cpp-python

```python
from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="NobodyWho/LFM2.5-8B-A1B-GGUF",
    filename="LFM2.5-8B-A1B-Q8_0-vendor-sampling.gguf",
)
```

## Model Specifications

- **Parameters:** 8B total / ≈1B active (MoE)
- **Context length:** 128,000 tokens
- **License:** LFM Open License v1.0
- **Base model:** LiquidAI/LFM2.5-8B-A1B
- **Architecture:** lfm2moe

## Licensing / Credits

Licensed under LFM Open License v1.0 (unchanged from upstream). All model credit belongs to
Liquid AI.