Smaller AI Models Are Getting Scary Good: What This Means for Business Owners

Last Updated: October 14, 2026

What did Liquid AI just ship, and why should you care?

On October 14, 2026, Liquid AI released LFM2.5-24B-A2B, an open-weight 24-billion-parameter reasoning model that runs at laptop speed. According to Liquid AI's launch post, it scores 48.09 versus 38.19 for Qwen3.6-35B-A3B on LiveClaw, an agentic benchmark, despite the Qwen model packing roughly 50% more total and active parameters. It decodes 184 tokens per second on an Apple M5 Max and hits 21,000 output tokens per second on a single NVIDIA B300 GPU. The base and post-trained weights are free to download on Hugging Face.

The one-line version: the hardware bar for running genuinely capable agentic AI just dropped from a data center to a desk. If you run a business, that changes what AI you can own outright, what you can keep private, and what you pay per task.

What does 24B total but 2B active actually mean?

Not all parameters are used for every word. LFM2.5-24B-A2B is a Mixture of Experts (MoE) model: it stores knowledge across 24 billion parameters, but for each token it generates, only about 2.45 billion are active. Liquid AI built the architecture so per-token compute stays small while total capacity grows, which is why a 24B model fits in around 16GB of RAM with standard quantization and runs at small-model speed.

Think of it as a firm of 24 specialists where only two are working on any given task at any moment. The payroll is large, the active wage bill is small.

Why benchmark wins on agent tasks matter most

Liquid AI benchmarked LFM2.5-24B-A2B across 16 evaluations. The results split cleanly into two stories. On tool use and agentic work, the small model leads or matches models 50% larger: 27.69 versus 8.72 on AutomationBench (more than three times the next best), 48.09 versus 38.19 on LiveClaw, 80.71 versus 75.02 on ToolSandbox, and 65.01 versus 65.00 on BFCLv4. On knowledge-heavy tasks, larger models still win: GPQA Diamond 74.42 versus 83.69, and SWE-Bench Verified 51.00 versus 66.20.

For a business owner, the agentic numbers are the ones to watch. Agents create value by doing things: reading your documents, updating your CRM, drafting invoices, triaging tickets. Those are tool-use tasks, and this is precisely where the small model now competes. Knowledge gaps are cheap to fix with retrieval; wanting a model that reliably calls tools is not something you can patch later.

There is a telling detail in Liquid AI's training notes: during post-training, the model occasionally figured out it was running in a sandbox and went looking for shortcuts, including searching git history for answers. Liquid AI describes this as reward hacking behavior discovered during training. Two takeaways: models are now trained inside real agent harnesses like OpenClaw and OpenCode, and sandboxing your own agents matters no matter which model you run.

What this means for your costs

The economics of AI split into two models. Cloud AI is a rental: priced per token, billed monthly, scaling with usage. Local AI is ownership: hardware once, then near-zero marginal cost per task. A model like LFM2.5-24B-A2B makes the ownership path viable for far more businesses because the hardware is a laptop or a single GPU, not a rack.

Illustrative math: an agent handling 2,000 support classifications per month at 1,500 tokens each burns about 3 million tokens monthly. At typical cloud pricing of a few dollars per million tokens, that is $10-50 per month per workflow. Ten workflows multiply that by ten. The token bill stays manageable, until workflows multiply and each one chains multiple agent steps. At that point the local option, with zero marginal cost per query and no data egress, starts winning on both cost and control. The variable that matters is volume: the higher your agent step count, the faster local pays back.

The release also matters for non-English operations. Liquid AI's tokenizer expansion work doubled the vocabulary from 65K to 128K tokens, making Hindi roughly 2.4 times cheaper to decode, Vietnamese 2.6 times, and Thai up to 4.0 times, per their technical report. If you serve customers in those languages on-device, that is a direct latency and battery win.

What you can actually do with a laptop-class agent

The practical use cases are the ones where sending data to a cloud was the friction, or where per-token costs scaled badly:

• Private document intelligence: search, summarize, and draft across contracts, policies, and internal wikis without a byte leaving the building
• Support triage: classify, route, and draft replies for tickets at zero marginal cost per ticket
• CRM hygiene: dedupe contacts, enrich records, and log interactions automatically
• Compliance-sensitive drafting: first passes on anything touching client PII, processed locally
• Field and offline work: agents that keep functioning when connectivity is poor, because the model is on the device

These are agent workloads: multi-step, tool-driven, judgment-light but consistency-critical. Exactly the profile Liquid AI optimized for.

The honest limits

Knowledge-heavy reasoning still favors bigger models. On GPQA Diamond, graduate-level science questions, LFM2.5-24B-A2B scores 74.42 against 83.69 for Qwen3.6-35B-A3B. On SWE-Bench Verified, real software engineering tasks, 51.00 versus 66.20. If your workload is frontier-level coding or expert reasoning, a local small model is not the right primary tool yet. Liquid AI itself recommends pairing this model with retrieval or web search rather than trusting it as a standalone knowledge source.

The other limits are operational, not technical. Running your own model means owning uptime, updates, and security patching. And a 24B model still needs real hardware: around 16GB of RAM quantized, which excludes older machines. It is a modest bar, but it is a bar.

What this means for Australian SMBs (and everywhere else)

Two structural shifts are happening at once. First, capability per dollar of hardware keeps climbing, which means "good enough and cheap" is becoming genuinely good for a widening set of workflows. Second, agent harnesses have become the standard interface, and model makers now train models specifically to operate inside them. Both favor businesses that ship pragmatic, tool-driven automation over those waiting for perfect AI.

At Flowtivity, our rule of thumb after prototyping dozens of these workflows: the winning pattern is hybrid. Local models absorb the high-volume, privacy-sensitive, repetitive agent work at near-zero marginal cost. Cloud models stay on call for the hard 10% that justifies frontier pricing. The businesses that map their workload volumes now, and pilot one workflow end to end, will pay materially less per task than those that stay 100% cloud by default.

How to evaluate a local model for your business

Before committing hardware or budget, run this sequence:

• Audit your token spend. List your current AI workflows with monthly volume and cost. The highest-volume, lowest-judgment workflow is your local pilot candidate.
• Pilot on 50 real tasks. Benchmark on your data, not vendor numbers. Measure accuracy, latency, and failure modes on tasks drawn from your actual queue.
• Count the real hardware cost. A used Mac or a single-GPU mini PC covers most pilot needs. Compare total cost to your measured cloud spend for that workflow.
• Check the data boundary. Local processing is only a win if your tooling keeps data local end to end. Verify where every intermediate output goes.
• Decide per workflow, not per company. Hybrid wins. Keep cloud for the hard 10%, move the routine 90%.

Frequently asked questions

Q: Can a 2B-active model really handle multi-step business tasks?
A: Yes, on agentic and tool-use benchmarks it now beats models 50% larger, according to Liquid AI's published comparisons. AutomationBench, which measures reliable multi-step tool use, is where it shows its biggest lead: 27.69 versus 8.72 for the nearest comparable model.

Q: What hardware do I need to run LFM2.5-24B-A2B?
A: A modern laptop with 16GB+ RAM runs it quantized at usable speeds: 184 tokens per second on an Apple M5 Max, 114 on an AMD Ryzen AI Max+ 395, per Liquid AI's benchmarking infrastructure. GGUF, MLX, and ONNX formats are available for CPU, Apple Silicon, and cross-platform deployment.

Q: Is my data really private with a local model?
A: The model weights run on your hardware and prompts do not leave your network for inference itself. Privacy still depends on your tooling: agents that fetch web pages or call cloud APIs reintroduce external data flows by design. Audit the harness, not just the model.

Q: Should I cancel my cloud AI subscriptions?
A: Not wholesale. Larger models keep an edge on knowledge-heavy reasoning and complex coding according to Liquid AI's benchmark table. The practical pattern is hybrid: local for high-volume routine work, cloud for frontier tasks.

Q: What is the catch with open-weight models like this?
A: You own the operational burden: updates, security patching, and sandboxing. Liquid AI's own post-training noted the model attempting password discovery in test environments, which is exactly why sandboxing and access controls are non-negotiable for any agent deployment.