LFM2.5-2.6B — Ollama Model (8 GB GPU) General-Purpose Agent

Model: oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU

DESCRIPTION

LFM2.5-2.6B is a next-generation agentic language model by Liquid AI, built on the hybrid LFM2 architecture — 22 double-gated short convolution (LIV) blocks combined with 8 Grouped-Query Attention (GQA) blocks. It is a pure reasoning model that always thinks before it answers (wrapping its reasoning in <think>/</think> tags), and was trained inside popular agentic harnesses to work reliably as an on-device agent. This Ollama packaging uses the official Q4_K_M GGUF quantization (~1.67 GB), bringing frontier-grade agentic performance to 8 GB GPUs and CPU-only laptops.

“Deploy Agents Everywhere.” — Liquid AI

The ideal on-device agent for: 8 GB laptops · 4-8 GB desktops · Edge servers · Raspberry Pi 5 · Agent harnesses (Hermes Agent, OpenClaw, Pi)

Key Features

• Best-in-class agent: competitive with models 4x larger on tool use, instruction following, and multi-step agentic tasks
• Native tool calling: Pythonic function calls (func(arg="value")) between <|tool_call_start|>/<|tool_call_end|> tokens
• Always thinks first: built-in reasoning (<think> tags) for planning and multi-step workflows
• 128K native context (configured to 32K for 8 GB, expandable to 128K at runtime)
• Tiny footprint: only ~1.67 GB at Q4_K_M — runs on any 8 GB GPU with ~5.5 GB of headroom
• Blazing fast: 220 tok/s on Apple M5 Max, 113 tok/s on AMD Ryzen, ~30 tok/s on a phone
• 16 languages: EN, AR, ZH, FR, DE, IT, JA, KO, PT, ES, VI, TH, ID, HI, RU, PL
• Trained on ~34T tokens: pre-training + context extension to 128K + agentic RL
• LFM Open License v1.0: royalty-free for non-commercial/research and small business (<$10M annual revenue)

Architecture

REQUIREMENTS

Platform support: - CPU-only — runs on any modern CPU (efficient hybrid architecture) - Any GPU (NVIDIA, AMD, Intel Arc, Apple Silicon) with 2 GB+ VRAM - Raspberry Pi 5 — confirmed by Liquid AI for on-device agents - Edge devices & phones — purpose-built for edge deployment

💡 Why 8 GB GPU? The model is only ~1.67 GB. An 8 GB GPU has ~5.5 GB of headroom — enough for large 128K contexts, KV cache, and concurrent requests.

QUICK START

1. Install Ollama

# macOS (Homebrew)
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows
# Download from https://ollama.com/download

2. Pull & Run

# Pull the model (downloads ~1.7 GB)
ollama pull oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU

# Run interactively
ollama run oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU

# Single prompt
ollama run oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU \
  "Plan a 3-step workflow to research and summarize the Q3 earnings of a public company."

USAGE

CLI

# Interactive chat
ollama run oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU

# With system prompt
ollama run oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU \
  --system "You are an agent. Plan and execute multi-step tasks using the available tools."

# Override context for long-context workflows
ollama run oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU \
  --num-ctx 131072

REST API

# Chat completion
curl -s http://127.0.0.1:11434/api/chat \
  -d '{
    "model": "oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU",
    "messages": [{"role": "user", "content": "What is the capital of France?"}]
  }'

# Multi-turn agentic task
curl -s http://127.0.0.1:11434/api/chat \
  -d '{
    "model": "oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU",
    "messages": [
      {"role": "system", "content": "You are a helpful agent. Think step by step."},
      {"role": "user", "content": "Extract all dates and amounts from: Invoice dated 2024-03-15 for $1,234.56"}
    ],
    "options": {
      "temperature": 0.1,
      "top_k": 50
    }
  }'

# Tool calling (prompt mode)
curl -s http://127.0.0.1:11434/api/chat \
  -d '{
    "model": "oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU",
    "messages": [
      {"role": "system", "content": "List of tools: [{\"name\": \"get_weather\", \"description\": \"Get the current weather for a city\", \"parameters\": {\"type\": \"object\", \"properties\": {\"location\": {\"type\": \"string\"}}, \"required\": [\"location\"]}}]"},
      {"role": "user", "content": "What is the weather in Boston?"}
    ]
  }'

# OpenAI-compatible endpoint
curl -s http://127.0.0.1:11434/v1/chat/completions \
  -d '{
    "model": "oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": false
  }'

Python (ollama library)

pip install ollama

import ollama

# Chat
response = ollama.chat(
    model='oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU',
    messages=[{'role': 'user', 'content': 'What is 2+2? Answer briefly.'}],
)
print(response.message.content)

JavaScript (ollama.js)

npm install ollama

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU',
  messages: [{ role: 'user', content: 'Plan a trip to Japan.' }],
})
console.log(response.message.content)

SAMPLING PARAMETERS

These are baked into the model via its Modelfile:

Equivalent llama.cpp Command

./llama-cli \
  -m LFM2.5-2.6B-Q4_K_M.gguf \
  -ngl 99 \
  -c 32768 \
  --temp 0.1 \
  --top-k 50 \
  --repeat-penalty 1.1 \
  -cnv

Override for More Creative Tasks

If you need less deterministic outputs (e.g., for creative text generation):

ollama run oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU \
  --temperature 0.7 --top-p 0.95

MEMORY & PERFORMANCE

VRAM Budget

Context Window Scaling (8 GB, with q8_0 KV cache)

Performance

TROUBLESHOOTING

Optimizations for 8 GB Systems

# Set KV cache to q8_0 for ~50% memory reduction
export OLLAMA_KV_CACHE_TYPE=q8_0
ollama serve

# In another terminal:
ollama run oamazonasgabriel/lfm2.5-2.6b:q4_k_m-8gbGPU

ALTERNATIVE MODELS

# Larger general-purpose MoE sibling (8 GB GPU)
ollama pull oamazonasgabriel/lfm2.5-8b-a1b:q4_k_m-8gbGPU

# Ultra-lightweight LFM2.5 (4-8 GB RAM)
ollama pull oamazonasgabriel/lfm2.5-230m:bf16-8gbRAM

# Tool-calling specialized (8-16 GB RAM)
ollama pull oamazonasgabriel/lfm2-1.2b-tool:f16-8gbRAM

# Larger coding model (needs 16 GB)
ollama pull oamazonasgabriel/qwen3.5-9b:q4-16gbGPU

# Larger MoE model (needs 24 GB)
ollama pull oamazonasgabriel/qwen3.6-35b-a3b:q4-24gbGPU

# Other quants of this model (Q8_0, Q6_K, etc.)
ollama pull hf.co/LiquidAI/LFM2.5-2.6B-GGUF:Q8_0

CREDITS

This Ollama packaging brings together work from multiple teams and individuals:

Citation

If you use this model in your work, please cite the original LFM2.5-2.6B blog post and the LFM2 technical report:

@article{liquidAI202626B,
  author  = {Liquid AI},
  title   = {LFM2.5-2.6B: Agents Everywhere},
  journal = {Liquid AI Blog},
  year    = {2026},
  note    = {www.liquid.ai/blog/lfm2-5-2-6b},
}

@article{liquidai2025lfm2,
  title={LFM2 Technical Report},
  author={Liquid AI},
  journal={arXiv preprint arXiv:2511.23404},
  year={2025}
}

Acknowledgments

• The llama.cpp project (github.com/ggml-org/llama.cpp) for the GGUF format and efficient inference engine that makes models like this accessible on commodity hardware.
• The Ollama team (ollama.com) for the seamless model distribution platform.
• Hugging Face (huggingface.co) for model hosting and the Hub ecosystem.

LINKS