---

license: other

license_name: lfm1.0

license_link: LICENSE

base_model: LiquidAI/LFM2.5-VL-450M

tags:

- gguf

- tool-calling

- vision

- liquid

- lfm2.5

pipeline_tag: image-text-to-text

---

LFM2.5-VL-450M GGUF — with vendor sampling metadata

GGUF builds of LiquidAI/LFM2.5-VL-450M

prepared for tool calling. Unlike most LFM GGUFs, this model ships a

complete native chat template (it already renders the tool_calls field),

so no template change was needed. Every file is the corresponding upstream

quant with bit-identical weight tensors and one metadata addition:

LiquidAI's recommended sampling settings embedded as general.sampling.*

(temp=0.1, min_p=0.15, penalty_repeat=1.05).

This is the smallest tool-capable LFM — a practical edge-device

tool-caller at 219-711 MB.

Model Capabilities

• Text generation — instruction-following chat model
• Tool calling — native LFM2 function-calling format, including multi-turn tool use
• Vision — understands and reasons about images (pair with the upstream mmproj file, see Getting Started)
• Long context — 128k tokens

Getting Started

> [!NOTE]

> Running this model with nobodywho requires the upcoming release

> (PR #564): its native chat template uses

> HuggingFace {% generation %} tags that nobodywho ≤ 1.4.0 cannot parse, and LFM tool calling

> ships in the same release. The files work in any other llama.cpp-based runtime; the original

> unmodified GGUFs live in the upstream

> LiquidAI/LFM2.5-VL-450M-GGUF repo.

Install NobodyWho:

pip install nobodywho

Run — the model is downloaded and cached automatically on first use:

from nobodywho import Chat

chat = Chat("huggingface:NobodyWho/LFM2.5-VL-450M-GGUF/LFM2.5-VL-450M-Q8_0-vendor-sampling.gguf")
response = chat.ask("What is the capital of Denmark?").completed()
print(response) # Copenhagen!

Tool calling

from nobodywho import Chat, tool

@tool(description="Gets the current weather for a city")
def get_weather(city: str) -> str:
    return f"It is sunny and 22°C in {city}."

chat = Chat(
    "huggingface:NobodyWho/LFM2.5-VL-450M-GGUF/LFM2.5-VL-450M-Q8_0-vendor-sampling.gguf",
    tools=[get_weather],
)
print(chat.ask("What is the weather in Paris?").completed())

Vision

This repo now hosts the language model and the matching projection

models (mmproj) — pass one as projection_model_path for image input.

Two precisions are available: mmproj-LFM2.5-VL-450m-F16.gguf and a smaller

mmproj-LFM2.5-VL-450m-Q8_0.gguf (either pairs with any model quant):

from nobodywho import Model, Chat, Prompt, Image, Text

model = Model(
    "huggingface:NobodyWho/LFM2.5-VL-450M-GGUF/LFM2.5-VL-450M-Q8_0-vendor-sampling.gguf",
    projection_model_path="huggingface:NobodyWho/LFM2.5-VL-450M-GGUF/mmproj-LFM2.5-VL-450m-F16.gguf",
)
chat = Chat(model, system_prompt="You are a helpful assistant.")

prompt = Prompt([
    Text("What do you see in this image?"),
    Image("./photo.png"),
])
response = chat.ask(prompt).completed()
print(response)

Files

Scores on NobodyWho's 14-test tool-calling suite. "metadata active" =

runtimes that read sampler defaults from the model file; "ignored" = runtimes

that don't (the embedded sampling consistently gains one test on this model).

| File | metadata active | metadata ignored |

|---|---|---|

| LFM2.5-VL-450M-Q8_0-vendor-sampling.gguf (379 MB) | 13/14 | 12/14 |

| LFM2.5-VL-450M-F16-vendor-sampling.gguf (711 MB) | 13/14 | 12/14 |

| LFM2.5-VL-450M-Q4_0-vendor-sampling.gguf (219 MB) | 12/14 | 11/14 |

The one consistent failure at the top configurations is a single test where

the model calls the right tool but garbles a string inside a tuple-typed

argument — verified stable across quants up to F16. No refusals, no crashes.

Use

Verified with NobodyWho

(see PR #564); works in

any llama.cpp-based runtime. Text-only use needs no mmproj; for vision, pair

with the mmproj-LFM2.5-VL-450m-* files hosted in this repo.

Model Details

| Property | Value |

|---|---|

| Parameters | 450M (354M language model + vision tower in the mmproj) |

| Context length | 128,000 tokens |

| License | LFM Open License v1.0 |

| Base model | LiquidAI/LFM2.5-VL-450M |

License

LFM Open License v1.0, unchanged from upstream — see LICENSE.

All credit for the model goes to Liquid AI.