TECH Oct 22, 2026 · by Skills Team

Liquid AI LFM2.5: Why Efficient AI Models Win Business

Liquid AI released LFM2.5-24B-A2B on October 13, 2026: 24B total parameters, only 2B active per token, beating bigger rivals on tool-use benchmarks. The split results prove buyers should evaluate models on their own workflows, not on parameter counts.

Liquid AI LFM2.5: Why Efficient AI Models Win Business

The Short Answer

For years, the AI industry treated one number as a shorthand for quality: the parameter count. Bigger was better, and every headline celebrated a larger model. Then Liquid AI LFM2.5 arrived on October 13, 2026, and challenged that habit. LFM2.5-24B-A2B carries 24 billion total parameters, but activates only about 2 billion of them for each…

Liquid AI LFM2.5: The Efficiency Shift Businesses Should Notice

For years, the AI industry treated one number as a shorthand for quality: the parameter count. Bigger was better, and every headline celebrated a larger model. Then Liquid AI LFM2.5 arrived on October 13, 2026, and challenged that habit. LFM2.5-24B-A2B carries 24 billion total parameters, but activates only about 2 billion of them for each token it generates. In Liquid AI's published testing, this leaner model beats Alibaba's Qwen3.6-35B-A3B on tool-use tasks, even though Qwen3.6 has roughly 50 percent more total and active parameters. The lesson for buyers is simple and overdue: choose the model that wins on the work you actually do, not the one with the biggest specification sheet.

What Liquid AI LFM2.5 Actually Is

Liquid AI LFM2.5-24B-A2B is an open-weight mixture-of-experts reasoning model released by Liquid AI on October 13, 2026. It is the largest and final model of the LFM2.5 family, a line Liquid AI spent the year building toward one goal: making capable AI run locally on consumer hardware instead of only inside distant data centers. Both the base checkpoint and the post-trained model are available as open weights on Hugging Face, so any team can download and test them independently.

The architecture keeps the LFM2-24B-A2B foundation: 40 layers with 64 experts per mixture-of-experts block and top-4 routing, meaning each token is handled by just 4 of the 64 experts. Two things changed from its predecessor: the context window grew from 32K to 131K tokens, and the vocabulary doubled from 65K to 128K tokens for better handling of non-Latin scripts. Pre-training used about 33 trillion tokens, up from about 17 trillion, followed by a 2-trillion-token mid-training phase and reinforcement learning inside agent harnesses such as OpenClaw, Hermes Agent, and OpenCode. The model supports 16 languages.

Why Liquid AI LFM2.5 Wins Where It Matters for Business

The headline result from the release is a split verdict. On tool-use and agentic benchmarks, Liquid AI reports that LFM2.5-24B-A2B leads or ties on five of eight tests against larger rivals including Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning. The standout numbers:

  • LiveClaw: 48.09 versus 38.19 for Qwen3.6-35B-A3B
  • AutomationBench: 27.69 versus 8.72 for Qwen3.6-35B-A3B, more than three times the next-best score in the comparison
  • Instruction following: leads on all three instruction-following benchmarks, including IFEval at 96.75

These are the benchmarks that mirror real business work. Tool use is what lets an AI model call your CRM, query your inventory database, draft a support ticket, or complete a multi-step workflow. If you are building agents that do work inside your company systems, a model that wins on these tests is more useful than one that wins elsewhere. This is also why AI development for business should start from task benchmarks rather than leaderboard rankings by size.

The Split Result Is the Real Insight

Here is the part most commentary misses. The same results table that shows LFM2.5 leading on agent tasks shows it trailing on others: Qwen3.6-35B-A3B scores 66.20 versus 51.00 on SWE-Bench Verified coding through OpenCode, and 83.69 versus 74.42 on GPQA Diamond, the knowledge-heavy benchmark. The bigger models keep the edge on deep coding and encyclopedic knowledge work.

This split is not a weakness of the release. It is the most important signal in it. No single model wins everything, and the era of ranking models by one number is over. The open-weight market is fragmenting into specialists: efficient models that excel at tool use and agents, larger models that excel at coding and knowledge work. For a business, that means the buying question changes from "which model is best?" to "which model is best at the workflow we are deploying?" Evaluate on your own pipeline, run your own tests, and pick the winner for your task.

Liquid AI LFM2.5 and the On-Device Advantage

Because only 2 billion parameters compute per token, each reasoning token is cheaper to run in raw compute terms, which is what makes on-device deployment realistic. Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, claiming the fastest of the compared models while using the least memory. On GPU, the company reports over 21,000 output tokens per second at high concurrency on a single NVIDIA B300. The model ships with day-one support in llama.cpp, MLX, vLLM, SGLang, and ONNX.

This matters beyond speed. Running agents on local hardware keeps sensitive data inside your own walls instead of sending it to a third-party API on every step. Teams in regulated industries and anyone handling customer data should take that seriously. Self-hosted AI models are becoming a practical enterprise option, and models like LFM2.5 are exactly what makes that shift possible.

The Bigger Trend: Efficiency Is the New Frontier

LFM2.5 did not land in a vacuum. It arrived in the middle of what one benchmark tracker counts as 22 confirmed model launches in October 2026 alone, including Meta's 816-billion-parameter Muse Spark Open two days later on October 15 and NVIDIA's Nemotron 3.5 Super VL. That contrast is telling: the same week produced both an 816B giant and a model that deliberately activates only 2B parameters per token.

Industry research keeps pointing the same direction. Small language models under 10 billion parameters are reported to deliver 80 to 90 percent of large-model capability on enterprise tasks while running entirely on consumer hardware. Local inference can run at a fraction of proprietary cloud API rates, and Gartner projects that task-specific small models will be used three times more than general large language models by 2027. Industry analysis also notes that inference now consumes roughly 80 percent of AI infrastructure budgets, which makes compute-per-token efficiency a genuine operational concern, not just a benchmark curiosity. If you want to understand the wider direction, read our take on on-device AI for business apps.

How Businesses Should Evaluate Models Like Liquid AI LFM2.5

The practical playbook for any team choosing AI in late 2026 looks like this:

  1. Define the workflow first. Is the model answering customer questions, calling tools, writing code, or reasoning over documents? The LFM2.5 story shows that the winner for one job may lose on another.
  2. Benchmark on your own data. Vendor benchmarks are a starting point, and Liquid AI's figures are its own published results. Because the weights are open, you can run your own tests before committing. Do that.
  3. Match the deployment to the task. Agent loops that run continuously benefit most from efficient models that can live on local hardware. Deep coding work may still favor the larger model. Many teams will end up with a mix, routing each step to the right tier. Our breakdown of Meta's Muse Spark Open shows the other end of the scale, and why giant models still have a role.
  4. Factor in data sensitivity. If a task runs on-device with an efficient model, customer data never leaves your building. That alone can decide the architecture for support, HR, and finance workflows. Custom AI knowledge bases are one example where self-contained deployment pays off.

What This Means for Skills Technology Solutions Clients

We build AI integrations around this exact principle. When a client asks for an AI assistant, an automation pipeline, or a customer support agent, we start with the workflow and select the model that scores best on that task, not the one with the largest parameter count. Open-weight efficient models like LFM2.5 are becoming core options in our AI and machine learning services, especially for agent-based workflows and on-device or self-hosted deployments where data control matters.

Frequently Asked Questions

What is Liquid AI LFM2.5-24B-A2B?

It is an open-weight mixture-of-experts reasoning model released by Liquid AI on October 13, 2026. It has 24 billion total parameters but activates only about 2 billion per token, and it is designed to run AI agents directly on devices rather than through a cloud API.

Does Liquid AI LFM2.5 beat bigger models?

On tool-use and agentic benchmarks, yes, according to Liquid AI's published results. It scores 48.09 on LiveClaw versus 38.19 for Alibaba's larger Qwen3.6-35B-A3B. But it trails the same competitor on coding (SWE-Bench Verified) and knowledge benchmarks, which proves the point that no model wins everywhere.

Can Liquid AI LFM2.5 run on a laptop?

Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395. It supports llama.cpp, MLX, vLLM, SGLang, and ONNX, but comfortable use appears to require a machine with substantial unified memory.

Why do smaller efficient models matter for businesses?

Efficient models can match or beat larger ones on specific tasks like tool use and agents while running on local hardware. That enables faster agent loops, keeps sensitive data on your own systems, and lets teams pick the right model per workflow instead of renting one giant model for everything.

How should my business choose between models?

Define your workflow first, benchmark candidate models on your own data, and deploy each task to the model that wins that task. The split results in the LFM2.5 release show that parameter counts are not a reliable ranking criterion.

Ready to put efficient AI to work in your business? We help companies choose, deploy, and integrate the right AI models for their actual workflows, from agent-based automation to self-hosted deployments. Contact Skills Technology Solutions today and tell us what you want AI to do for you.

#Liquid AILFM2.5efficient AI modelsAI agentson-device AIAI for business