LFM2.5-1.2B-Instruct GGUF — with tool_calls chat-template fix

GGUF builds of LiquidAI/LFM2.5-1.2B-Instruct prepared for tool calling. Every file is the corresponding upstream quant with bit-identical weight tensors; the embedded chat template (tokenizer.chat_template) is extended to render the tool_calls field of assistant messages, and the files marked in the table additionally embed LiquidAI's recommended sampling as general.sampling.* metadata.

Model Capabilities

• Text generation — instruction-following chat model
• Tool calling — native LFM2 function-calling format; multi-turn tool use works thanks to the template fix in this repo
• Long context — 128k tokens

Getting Started

Install NobodyWho:

pip install nobodywho

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

from nobodywho import Chat

chat = Chat("huggingface:NobodyWho/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-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/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-Q8_0-vendor-sampling.gguf",
    tools=[get_weather],
)
print(chat.ask("What is the weather in Paris?").completed())

[!NOTE] Tool calling with LFM models ships in the upcoming nobodywho release (PR #564). These files also work in any other llama.cpp-based runtime; the original unmodified GGUFs live in the upstream LiquidAI/LFM2.5-1.2B-Instruct-GGUF repo.

Why

The upstream template renders only message.content. Runtimes that store tool calls in the structured tool_calls field (the HF "unified tool use" convention, used by NobodyWho and OpenAI-style APIs) re-render assistant tool-call turns as empty turns, so the model never sees its own previous calls — causing re-issued tool calls and degraded multi-turn tool use.

The fixed template renders them in the model's native markup:

<|tool_call_start|>[get_weather(city="Paris")]<|tool_call_end|>

Files

Sampling notes

The Q4_0 and Q8_0 files embed LiquidAI's recommended sampling as general.sampling.* metadata, taken from the vendor's LEAP deployment config: temperature 0.3, min_p 0.15, repetition_penalty 1.05. The F16 deliberately ships without sampling metadata: at full precision, embedding those values drops the bash-writing test (13/14 vs 14/14 with default sampling), so runtimes fall back to their own defaults.

The embedded values were previously temp 0.1, top_k 50 (the vendor's model-card prose, which conflicts with its LEAP config) and have been corrected to the LEAP values above. All scores re-verified against these files with the corrected sampler: Q8_0 14/14, F16 14/14 (sampling-free), F16 + vendor sampling 13/14, Q4_0 12/14 (double-calls two tests).

Use

Verified against NobodyWho's 14-test tool-calling suite (single and multi-call, nested arguments, multi-turn) — see PR #564. Works as a drop-in replacement for the upstream Q8_0 file in any llama.cpp-based runtime.

Model Details

License

LFM Open License v1.0, unchanged from upstream — see LICENSE. All credit for the model goes to Liquid AI.