LFM2.5-2.6B: How to Run Liquid AI's On-Device Agent Locally

What You'll Learn

• What makes LFM2.5-2.6B different from other small open-weight models
• Real benchmark scores against Gemma 4 and Qwen3.5 models up to 4x its size
• Exact system requirements, memory footprint and hardware options
• Step-by-step install instructions for Mac, Windows and Linux using llama.cpp, LM Studio and MLX

LFM2.5-2.6B, released by Liquid AI on August 4, 2026, is an open-weight on-device agent model built specifically for tool use, instruction following and multi-step agentic tasks. Where most small models force developers to trade capability for size, Liquid AI claims this 2.6B checkpoint stays competitive with models nearly four times larger while decoding at 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen CPU, all inside a 2.5 GB memory footprint. The timing matters: agentic workloads are moving off cloud APIs toward local inference for privacy and near-zero cost, and LFM2.5-2.6B is one of the first models trained inside real agent harnesses like Hermes Agent and OpenClaw during reinforcement learning.

What Is LFM2.5-2.6B? A 2.6B Agent Built for Edge Devices

LFM2.5-2.6B is the largest checkpoint in Liquid AI's LFM2.5 family of hybrid models designed for on-device deployment. It builds on the LFM2 architecture, which interleaves 22 double-gated short convolution blocks with 8 grouped-query attention (GQA) blocks across 30 layers. The model totals 2.69 billion parameters, uses a 128,000-token vocabulary and supports a 131,072-token (128K) context window, trained on roughly 34 trillion tokens.

The model ships in two versions: LFM2.5-2.6B-Base, a pre-trained checkpoint for fine-tuning, and LFM2.5-2.6B, the post-trained agentic release. Post-training runs through four stages: supervised fine-tuning, per-domain teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning with GRPO. That final stage trains the model inside real agent harnesses, including Hermes Agent and OpenClaw, exposing it to their tools, system prompts and interaction patterns. The result is a pure reasoning model that always thinks before answering, adding a think tag directly in its ChatML-style chat template.

Liquid AI recommends the model for agentic workloads, tool use, data extraction, retrieval-augmented generation and long-context workflows, while noting it is not the right choice for agentic coding or knowledge-heavy tasks. Developers pairing it with a capable harness can build private, always-on agents that burn through millions of tokens at no marginal cost.

LFM2.5-2.6B Benchmarks: Beating Models 4x Larger

Liquid AI evaluated LFM2.5-2.6B against relevant sub-10B models across STEM, instruction following, tool use and agentic workflows. Despite being the smallest model in the comparison, it leads on every instruction-following benchmark and nearly every tool-use benchmark, trailing only Qwen3.5-9B on BFCLv4.

On agentic tasks it outperforms both Gemma models across the board and trades closely with the Qwen models. Coding is the one area where the larger models keep a clear edge, which is why Liquid AI positions LFM2.5-2.6B for high-volume agentic workloads on edge devices rather than for code generation. For a deeper look at how small open models fit into production stacks, our Cohere Aya Expanse and Tiny Aya guide covers a similar class of efficient models.

System Requirements: Run It on a Mac, PC or Raspberry Pi

The headline spec is the 2.5 GB memory footprint. Because LFM2.5-2.6B uses the efficient LFM2 hybrid architecture, it runs comfortably on consumer hardware without a discrete GPU. Liquid AI's published numbers: 220 tokens per second on an Apple M5 Max, 113 tokens per second on an AMD Ryzen AI Max+ 395 CPU, and roughly 30 tokens per second on a phone, which is enough to run a responsive agent. VentureBeat's testing reported the model ran 3.7 times faster than DeepSeek-V4-Flash on comparable hardware and confirmed it works on devices as small as a Raspberry Pi.

For GPU serving, a single NVIDIA H100 can sustain almost 15,000 output tokens per second at high concurrency, or about 1.3 billion tokens per day. On the desktop, a MacBook with Apple Silicon or a Windows PC with 8 GB of RAM is plenty for interactive use, and GGUF quantization reduces the footprint further for CPU-only machines. The model ships in native, GGUF, ONNX and MLX formats, with day-one support for llama.cpp, MLX, vLLM, SGLang and LM Studio.

Before scaling out, teams should also plan the security layer around local agents. Our OWASP Top 10 for LLM Applications guide covers the attack surface that any agentic deployment inherits, including tool-injection and prompt-injection risks.

How to Install LFM2.5-2.6B Locally (Mac, Windows, Linux)

LFM2.5-2.6B is available on Hugging Face as LiquidAI/LFM2.5-2.6B, and setting it up takes a few minutes on any platform. The easiest route for desktop users is LM Studio: open the app, search for LiquidAI/LFM2.5-2.6B in the model browser, download the GGUF quant, and load it. It works out of the box on macOS, Ubuntu and Windows because LM Studio is built on llama.cpp.

For developers who prefer the command line, the official Hugging Face model card provides a Transformers quick start. Install transformers 5.0.0 or newer, then load the model with AutoModelForCausalLM.from_pretrained('LiquidAI/LFM2.5-2.6B', device_map='auto') in bfloat16. Generation parameters recommended by Liquid AI are temperature 0.1, top_k 50 and repetition_penalty 1.1.

For an OpenAI-compatible serving endpoint, run vllm serve LiquidAI/LFM2.5-2.6B after pip install vllm, then point any client at http://localhost:8000/v1/chat/completions. The same pattern works with SGLang, and Apple Silicon users can use the MLX format for optimized inference. Once the endpoint is live, point your agent harness at it: Liquid AI says LFM2.5-2.6B works out of the box with Hermes Agent, OpenClaw and Pi. If you are building the routing layer yourself, our multi-model API fallback guide shows how to add zero-downtime failover between local and cloud models.

LFM2.5-2.6B vs Qwen3.5-9B and Gemma 4: Where It Wins

The most useful comparison is against Qwen3.5-4B, Qwen3.5-9B, gemma-4-E2B-it and gemma-4-E4B-it, the models Liquid AI benchmarked directly. LFM2.5-2.6B beats every Gemma checkpoint on tool use and instruction following, and on ToolSandbox it scores 77.83, ahead of Qwen3.5-9B's 76.44. On Multi-IF it reaches 80.07 versus 62.55 for Qwen3.5-9B, a significant gap in multi-turn instruction following.

The trade-offs are clear. Qwen3.5-9B wins on math (AIME25: 56.07 vs 51.87) and coding (LiveCodeBench: 69.86 vs 59.41), and its larger 9.7B parameter count gives it more raw knowledge capacity. But LFM2.5-2.6B delivers that agentic performance in a fraction of the memory and at much higher token rates, which changes the economics for high-volume, privacy-sensitive workloads. For budget-conscious builders, the free local inference also compares well against paid APIs in our AI model pricing comparison.

Pricing and License: Is LFM2.5-2.6B Free?

Yes. LFM2.5-2.6B is open-weight and free to download, run and fine-tune, including for commercial use, directly from Hugging Face. There is no per-token API charge when you run it locally, which is the core value proposition: removing the marginal cost of tokens lets developers parallelize agents across local hardware and run background tasks around the clock. The model has been climbing the OpenRouter leaderboard since release, according to VentureBeat, and third-party quantized versions are already available for Ollama. The broader LFM2.5 family follows the same open-weight model, with checkpoints from 230M up to 8B parameters.

Should You Run LFM2.5-2.6B?

LFM2.5-2.6B is the strongest open-weight option we have seen for on-device agentic workloads in the sub-3B class. It is not a coding model and it will not beat frontier models on knowledge-heavy tasks, but for tool use, instruction following, RAG pipelines and private local agents, it outperforms models up to four times its size while running in 2.5 GB of memory. Liquid AI's own guidance is to try it for high-volume agentic tasks where speed, privacy and cost matter most. The agent ecosystem is moving toward local execution, and models like this one, alongside the personal AI agents race, are why.