Workflows/LFM2.5 2.6B local agent setup

Agents / WORKFLOW

LFM2.5 2.6B local agent setup

Metadata for building a local tool-using agent with LFM2.5 2.6B using Transformers or llama.cpp across Windows, Linux, and macOS.

EstimatedSource checked 9/28/2026

LOCALRENTED GPU

Freshness 74 out of 100. Repository age: 79 days old, Jul 28, 2026. Measures age, not quality.

Popularity 70 out of 100. 114,916 downloads in 30 days; 831 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.

Check the date source

Popularity · adoption

Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

Before you begin

Hardware

Machine	LFM2.5-2.6B	LFM2.5-2.6B-GGUF	LFM2.5-2.6B-Heretic-Abliterated-GGUF
NVIDIA GPU (8 GB VRAM)	Transformers bfloat16 weights ~5.4GB; 8GB VRAM tight, use low context or CPU offload.	llama.cpp CUDA: Q4/QAD Q4_0 ~2GB + 2-4GB overhead fits 8GB.	llama.cpp CUDA: Q4_K_M 1.67GB + 2-4GB context overhead fits 8GB.
AMD GPU (8 GB VRAM)	Transformers ROCm on Linux: weights ~5.4GB; 8GB VRAM tight; CPU offload may help.	llama.cpp ROCm/Vulkan: Q4/QAD Q4_0 ~2GB + 2-4GB overhead fits 8GB.	llama.cpp ROCm/Vulkan: Q4_K_M 1.67GB + 2-4GB overhead fits 8GB.
Intel Arc GPU (8 GB VRAM)	Transformers on Intel Arc is not covered in this guide; use liquidai-lfm2-5-2-6b-gguf via Vulkan.	llama.cpp Vulkan: Q4/QAD Q4_0 ~2GB + 2-4GB overhead fits 8GB.	llama.cpp Vulkan: Q4_K_M 1.67GB + 2-4GB overhead fits 8GB.
Apple Silicon (16 GB unified memory)	Transformers MPS/CPU: weights ~5.4GB; 16GB unified likely fits with low context.	llama.cpp Metal: Q4/QAD Q4_0 ~2GB + 2-4GB overhead fits 16GB unified.	llama.cpp Metal: Q4_K_M 1.67GB + 2-4GB overhead fits 16GB unified.
CPU (16 GB system RAM)	Transformers CPU: weights ~5.4GB; 16GB system RAM fits with context headroom.	llama.cpp CPU: Q4/QAD Q4_0 ~2GB + 2-4GB overhead fits 16GB RAM.	llama.cpp CPU: Q4_K_M 1.67GB + 2-4GB overhead fits 16GB RAM.

Context window

Machine	LFM2.5-2.6B	LFM2.5-2.6B-GGUF	LFM2.5-2.6B-Heretic-Abliterated-GGUF
NVIDIA GPU (8 GB VRAM)	Start 2048 tokens; publisher limit 131k; 5.4GB weights leave little 8GB headroom.	Start 4096 tokens; publisher limit 131k; Q4 ~2GB fits 8GB with headroom.	Start 4096 tokens; publisher limit 131k; 1.67GB weights fit 8GB with headroom.
AMD GPU (8 GB VRAM)	Start 2048 tokens on Linux ROCm; 5.4GB weights leave little 8GB headroom.	Start 4096 tokens; publisher limit 131k; Q4 ~2GB fits 8GB with headroom.	Start 4096 tokens; publisher limit 131k; ROCm/Vulkan fits 8GB with headroom.
Intel Arc GPU (8 GB VRAM)	No Intel Arc Transformers route; use liquidai-lfm2-5-2-6b-gguf at 4096 tokens.	Start 4096 tokens; publisher limit 131k; Vulkan fits 8GB with headroom.	Start 4096 tokens; publisher limit 131k; Vulkan fits 8GB with headroom.
Apple Silicon (16 GB unified memory)	Start 4096 tokens; publisher limit 131k; MPS/CPU fits 16GB unified with headroom.	Start 8192 tokens; publisher limit 131k; Metal fits 16GB unified with headroom.	Start 8192 tokens; publisher limit 131k; Metal fits 16GB unified with headroom.
CPU (16 GB system RAM)	Start 4096 tokens; publisher limit 131k; CPU RAM fits 16GB with headroom.	Start 4096 tokens; publisher limit 131k; CPU RAM fits 16GB with headroom.	Start 4096 tokens; publisher limit 131k; CPU RAM fits 16GB with headroom.

Metadata onlySOURCE-LINKED SETUP

Set up LFM2.5 2.6B for local agentic tool calling

Use the publisher's native Transformers weights, the publisher's GGUF quantizations, or a community uncensored GGUF variant with llama.cpp. Verify a complete tool-call round-trip.

Choose a package Run its commands Check the result

BEFORE YOU BEGIN

Choose your package

Select the weights you want. The instructions, downloads and settings change with your choice.

The publisher has not provided a complete memory target for this task. Check its hardware guidance before starting.

CURRENT SETUP

LFM2.5-2.6B-GGUF · official

Estimate: add 2–4 GB RAM overhead for context. Works on CPU or GPU via llama.cpp backends.

Freshness 75 out of 100. Repository age: 75 days old, Aug 1, 2026. Measures age, not quality.

Popularity 78 out of 100. 904,724 downloads in 30 days; 379 lifetime likes. Checked Oct 15, 2026. Measures adoption, not quality.

What do these scores mean?

Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.

Freshness · release age

Check the date source

Popularity · adoption

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source

Scores as of Oct 15, 2026. Statistics checked Oct 15, 2026. Statistics are marked for refresh after 72 hours.Full scoring method

View weight source

THE WORKFLOW

Set up LFM2.5-2.6B-GGUF on your machine

Choose where you want to run it. The instructions below update to match your selection.

CHOOSE YOUR SETUPMac

Use llama.cpp with Metal or CPU backend. Prebuilt binaries or build from source.

Prerequisites and working folder (Terminal)

You need a Mac with Apple Silicon (Metal) or Intel (CPU), 16 GB unified memory recommended, and about 7 GB free disk for the BF16 GGUF plus runtime. The model maximum context is 131,072 tokens (config.json max_position_embeddings), but this guide starts with an active runtime context of 8192 tokens. Open Terminal and create the working folder that will hold the runtime binary, the GGUF and the publisher's chat template.

work="$HOME/LFM2.5-2.6B"
mkdir -p "$work"
cd "$work"
pwd
sw_vers
uname -m
sysctl -n hw.memsize

Install the llama.cpp prebuilt macOS binary

Use the official prebuilt macOS archive from the llama.cpp releases page. For Apple Silicon choose the artifact whose name ends in -bin-macos-arm64.zip; for Intel choose -bin-macos-x64.zip. Move the downloaded archive into your working folder, unzip it there, and confirm the binary runs. The unzipped llama-cli and llama-server live directly in the working folder. Keep this Terminal available for the rest of the guide.

llama.cpp releases page (download the macOS archive) llama.cpp build documentation

work="$HOME/LFM2.5-2.6B"
ls -lh "$HOME/Downloads" | grep -i llama
unzip -o "$HOME/Downloads/llama-"*-bin-macos-*.zip -d "$work"
ls -lh "$work"
"$work/llama-cli" --version

Download the exact publisher GGUF and its chat template

Download the exact publisher file LFM2.5-2.6B-BF16.gguf from the LiquidAI/LFM2.5-2.6B-GGUF repository at the pinned revision e7caca5d835a3901a8e0d63e94009429bafafdfc, plus the qad/chat_template.jinja that renders the LFM2.5 chat and tool-call format and the qad/generation_config.json that carries the publisher sampling defaults (temperature 0.1, top_k 50, repetition penalty 1.1). These three files are all you need for a llama.cpp agent endpoint.

Put these files in $HOME/LFM2.5-2.6B/3 required files

$HOME/LFM2.5-2.6B/

LFM2.5-2.6B-BF16.ggufExact publisher BF16 GGUF weight file for llama.cppGet file

$HOME/LFM2.5-2.6B/qad/

chat_template.jinjaPublisher chat template that renders the LFM2.5 chat and tool-call formatGet file generation_config.jsonPublisher generation defaults (temperature, top-k, repetition penalty)Get file

LFM2.5-2.6B-GGUF repository Liquid AI tool use documentation

work="$HOME/LFM2.5-2.6B"
mkdir -p "$work/qad"
cd "$work"
curl -L -o LFM2.5-2.6B-BF16.gguf "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF/resolve/e7caca5d835a3901a8e0d63e94009429bafafdfc/LFM2.5-2.6B-BF16.gguf"
curl -L -o qad/chat_template.jinja "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF/resolve/e7caca5d835a3901a8e0d63e94009429bafafdfc/qad/chat_template.jinja"
curl -L -o qad/generation_config.json "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF/resolve/e7caca5d835a3901a8e0d63e94009429bafafdfc/qad/generation_config.json"
ls -lh "$work" "$work/qad"

Launch the local OpenAI-compatible server (keep this Terminal open)

Start llama-server against the downloaded GGUF. Metal is used automatically on Apple Silicon, so no extra backend flag is required. Set --alias LFM2.5-2.6B-BF16 so /v1/models returns a simple, predictable model id for the client. --ctx-size 8192 is the active runtime context for this guide; the model maximum is 131,072 tokens. --jinja together with the publisher's qad/chat_template.jinja enables the LFM2.5 chat and tool-call format. The publisher sampling defaults 0.1 / 50 / 1.1 are passed explicitly. This process stays running, so leave the Terminal open and use a second Terminal for requests and the client.

llama.cpp server documentation Liquid AI tool use documentation

work="$HOME/LFM2.5-2.6B"
cd "$work"
"$work/llama-server" -m LFM2.5-2.6B-BF16.gguf --host 127.0.0.1 --port 8080 --ctx-size 8192 --alias LFM2.5-2.6B-BF16 --jinja --chat-template-file qad/chat_template.jinja --temp 0.1 --top-k 50 --repeat-penalty 1.1

Verify the endpoint and run the known-answer smoke test

Open a second Terminal. Confirm the server is up and read the exact served model id from /v1/models; with --alias set above it is LFM2.5-2.6B-BF16. Also run the publisher-documented chat command against the same GGUF with the publisher's --conversation flag and sampling defaults. When the REPL opens, enter the prompt 'What is the chemical symbol for gold?' and confirm the reply contains Au, then ask a follow-up in the same session to prove a real conversation. Exit with Ctrl+C.

llama.cpp CLI documentation llama.cpp server documentation

curl -s http://127.0.0.1:8080/v1/models
work="$HOME/LFM2.5-2.6B"
cd "$work"
"$work/llama-cli" -m LFM2.5-2.6B-BF16.gguf --conversation --jinja --chat-template-file qad/chat_template.jinja --temp 0.1 --top-k 50 --repeat-penalty 1.1

Install Visual Studio Code and the Cline extension

Download the macOS build of Visual Studio Code from the official site and drag Visual Studio Code.app into /Applications. Launch VS Code, press Cmd+Shift+X to open the Extensions panel, search for 'Cline' by publisher saoudrizwan, and install it. Cline's OpenAI-compatible provider is documented at the link below and is configured in the next step.

Visual Studio Code downloads Cline OpenAI-compatible provider documentation

# 1. Download the macOS .dmg from https://code.visualstudio.com/download and drag Visual Studio Code.app into /Applications
# 2. Open VS Code, press Cmd+Shift+X, search 'Cline' (publisher: saoudrizwan) and click Install
# 3. Reload the VS Code window when prompted

Point Cline at the local llama-server

In VS Code, open the Cline settings with the gear icon and set API Provider to 'OpenAI Compatible'. Set Base URL to http://127.0.0.1:8080/v1, API Key to any non-empty string such as sk-noop (the local server does not authenticate), and Model ID to LFM2.5-2.6B-BF16, which is the exact id returned by /v1/models because the server was started with --alias LFM2.5-2.6B-BF16. Set Context Window to 8192 so it matches --ctx-size, set Max Output Tokens to 512, and enable Computer Use / tool calling so Cline may issue tool calls. Save.

Cline OpenAI-compatible provider documentation

# In VS Code: Cline settings (gear icon)
# API Provider:   OpenAI Compatible
# Base URL:       http://127.0.0.1:8080/v1
# API Key:        sk-noop
# Model ID:       LFM2.5-2.6B-BF16
# Context Window: 8192   |   Max Output Tokens: 512
# Computer Use / tool calling: enabled

Create a disposable project and the input data

Back in the second Terminal, create a small scratch workspace with one JSON data file. The agent task in the next step reads this file and produces a derived report, which forces a read tool call and a write tool call against the same data. Open the folder in VS Code with File > Open Folder so Cline treats it as the workspace root.

Liquid AI tool use documentation

proj="$HOME/LFM2.5-2.6B-AgentProject"
mkdir -p "$proj/data"
cd "$proj"
cat > data/candidates.json <<'JSON'
[
  {"candidate_id": "12345", "status": "Interview Scheduled", "position": "Clinical Research Associate", "date": "2023-11-20"},
  {"candidate_id": "67890", "status": "Screening", "position": "Data Analyst", "date": "2023-12-01"}
]
JSON
cat data/candidates.json

Bounded agent task through Cline (real tool round-trip)

With the folder open in VS Code, switch to the Cline chat panel and enter one precise request. Cline emits read and write tool calls; approve each one when it prompts you (leave auto-approve off for writes). The llama-server Terminal logs the matching /v1/chat/completions requests, which shows the model driving the tool calls rather than answering from memory.

Liquid AI tool use documentation Cline OpenAI-compatible provider documentation

# In VS Code: File > Open Folder... and choose $HOME/LFM2.5-2.6B-AgentProject
# In the Cline chat, enter this exact request:
#   Use your file tools to read data/candidates.json. Then write summary.md containing:
#   (1) the number of candidates, (2) the candidate_id of each candidate, and
#   (3) the status of candidate 12345. Use the read and write tools; do not answer from memory.
# Approve each tool call when Cline asks. Do not enable auto-approve for writes.

Inspect the tool round-trip, the output and context settings

After approving the tool calls, cat summary.md in Terminal and confirm it reports 2 candidates, ids 12345 and 67890, and that candidate 12345 is in the 'Interview Scheduled' stage. In the Cline chat panel you should see the read tool call, the tool result, the write tool call and the final natural-language answer. The llama-server Terminal shows the matching request/response cycle. The 131,072-token value is the model maximum; the active runtime context is whatever you pass to --ctx-size. If the Mac is low on unified memory, restart llama-server with a smaller --ctx-size such as 4096 and set Cline's Context Window to the same number so prompt truncation matches.

llama.cpp server documentation LFM2.5-2.6B-GGUF repository

proj="$HOME/LFM2.5-2.6B-AgentProject"
cat "$proj/summary.md"
# Expected content: 2 candidates; ids 12345 and 67890; candidate 12345 status = Interview Scheduled.
# Also confirm the Cline chat panel shows the read and write tool calls and their results.
work="$HOME/LFM2.5-2.6B"
cd "$work"
# If memory is tight, restart the server with a smaller active context, e.g.:
# "$work/llama-server" -m LFM2.5-2.6B-BF16.gguf --host 127.0.0.1 --port 8080 --ctx-size 4096 --alias LFM2.5-2.6B-BF16 --jinja --chat-template-file qad/chat_template.jinja

SUCCESS CHECK · LFM2.5-2.6B-GGUF

Agentic tool round-trip

The model must load and generate text via the chosen runtime. A complete tool call round-trip must succeed: user query, model emits a tool call, tool returns a result, model produces a final answer using that result. The output should follow the publisher's tool use format.

https://huggingface.co/LiquidAI/LFM2.5-2.6BOpens on your machine

This metadata outlines the setup. Detailed hardware tables, context matrices, and executable instructions are covered in later stages.

Back to setup steps

When it doesn’t go to plan

Model fails to load due to dtype or architecture not supported by installed Transformers version.

Update Transformers to a version that supports Lfm2ForCausalLM (transformers_version 5.2.0 in config). Install via `pip install transformers[torch] --upgrade` and ensure PyTorch 2.5+ is present.

Tool calls are not parsed or emitted correctly during agent interaction.

Ensure the chat template from the model repository (chat_template.jinja) is applied via `apply_chat_template`. The template defines special tokens for tool call start/end. Confirm your client or framework uses this exact template.

REFERENCE LIBRARY

Sources and files

Original instructions, model files, and compatibility notes behind this setup.

6 SOURCES

01LFM2.5 2.6B model cardhuggingface.co 02LFM2.5 2.6B GGUFhuggingface.co 03LFM2.5 2.6B Heretic Abliterated GGUFhuggingface.co 04Transformers installationhuggingface.co 05llama.cpp repositorygithub.com 06llama.cpp server documentationraw.githubusercontent.com

The model behind this workflow

LFM2.5 2.6B

Keep exploring

Text

SETUP GUIDE

Qwen3.5 · 9B private chat setup

Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.

Freshness 41 out of 100. Release age: 230 days old, Feb 27, 2026. Measures age, not quality.

Build a private coding assistant with Qwen3.8 27B

Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.

Freshness 76 out of 100. Repository age: 71 days old, Aug 5, 2026. Measures age, not quality.

Arena Open Source Web coding: rank 10 of 55. Ranked model: qwen3.8-27b. Original model ranking, not a benchmark of this package. Export dated Oct 13, 2026. Source checked Oct 15, 2026. Human preference ranking, not hardware suitability.+3

SOFTWAREOllama + Visual Studio Code + Cline + Python

Create images locally with FLUX.2 Klein

Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.

Freshness 35 out of 100. Release age: 273 days old, Jan 15, 2026. Measures age, not quality.

SOFTWAREComfyUI

LOCALRENTED GPU

Verified source