Your AI Shouldn’t Stop Working When the Wi-Fi Drops

You’ve probably tried a few AI tools for your business. A chatbot for customer questions here, a summary generator for long reports there. It works — until the internet connection stutters, or the service changes its terms, or you stop to ask yourself: should our customer data really be sitting on a server on the other side of the planet?

That question matters more in Malaysia than most software vendors admit. Your customer records, supplier invoices, and internal pricing notes cross borders every time you paste them into an online AI. And in half the country — think warehouse counters in Shah Alam, clinics in Kuching, or retail shops in Johor Bahru — the connection itself is the weak point.

You now have a third option. Liquid AI released LFM2.5-2.6B, a small model that runs entirely on your own device. It plans multi-step tasks, calls tools, and works through long documents without sending a single prompt to a third-party API.

TL;DR: LFM2.5-2.6B is a 2.69B-parameter model that runs locally on phones, laptops, PCs, and robots. It handles tool calling and instruction-following at a level competitive with models nearly four times its size, reads up to 131,072 tokens of context in one go, and ships its weights openly for fine-tuning. If you’ve been holding back from AI because of data privacy or patchy internet, this is the moment to look again.

What “On-Device Agentic” Actually Means

Let’s unpack the jargon. An agentic model doesn’t just answer a single question. It works through a task the way an employee would: read the input, decide a sequence of steps, use a tool — check a database, update a spreadsheet, send a notification — and keep going until it finishes. On-device means every step happens locally, on hardware you own. The model covers 16 languages and is text-only, which sets the right expectation: it reads and writes, but doesn’t analyse images.

The specifications matter here. The model has 2.69 billion parameters across 30 layers, and was pre-trained on roughly 34 trillion tokens. The headline feature for you is the context window: 131,072 tokens means the model can hold an entire 200-page business agreement, or two months of supplier invoices, in working memory at once — no more pasting a document in four chunks because the chatbot keeps “forgetting” the first part.

Two versions are available: LFM2.5-2.6B for agentic work, and the Base checkpoint if you plan to fine-tune it with your own data. Both are open-weight and ship in GGUF, MLX, and ONNX formats with day-one support in llama.cpp, vLLM, and LM Studio. In plain language: if you or your IT vendor can run open-source software, you can run this model on hardware you already own.

One word of caution: this is not a coding model. The benchmarks show LiveCodeBenchv6 scores of 59.41 versus 69.86 for a much larger Qwen model. Your software development team should keep their existing tools. But for documents, forms, data extraction, and business automation, this model is squarely in the conversation.

How This Applies to Malaysian SMEs

Clinic or law firm: keep patient and client data inside your walls. Many Malaysian clinics still run booking and patient intake through WhatsApp, manual forms, and one aging desktop system. The moment you use an online AI to summarise a referral letter, that letter has left your building. With an on-device model running on a machine already in your office, extraction and summarisation happen locally — which sits comfortably with the Personal Data Protection Act’s expectation of responsible handling, since there’s simply no foreign server storing a copy. The 128K context window handles an entire medical report, not just a paragraph.

Warehouse or logistics business in Shah Alam, Penang, or Sarawak: Your counter staff don’t care about model architectures. They care about getting answers while the internet flickers. A local model in GGUF format, run through a tool like LM Studio on a standard office PC, can triage delivery orders, answer “which orders from supplier A are still open?”, and summarise a stack of scanned documents into a single report. Because the model supports tool calling, it can update a tracking table directly rather than just telling you what it found. The release notes specifically highlight form and invoice extraction, offline document triage, and robotics command parsing as recommended workloads — relevant if you deal with light industrial equipment or warehouse automation.

Retail or e-commerce teams: talk to your own data without leaking it. Many Malaysian retailers rely on third-party chatbot platforms, which means your product descriptions, stock availability, and return policies flow through a vendor’s pipe. An on-device assistant can read your product catalogue and your return policy, then answer customers accurately. Liquid AI lists e-commerce and financial services among its target industries, and the 16-language coverage means Bahasa Malaysia, English, Mandarin, and Tamil can coexist in the same workflow — a practical advantage for a country where staff and customers mix languages casually.

There’s also the industrial angle. The model runs on robots and embedded systems, and Liquid AI targets automotive, industrial robotics, and consumer electronics — all sectors where Malaysia has real manufacturing presence. For an SME doing assembly or component supply, a local model that parses voice commands or maintenance logs without phoning home is genuinely useful on a factory floor where the Wi-Fi is weak by design.

Where It Beats Bigger Models (and Where It Doesn’t)

Across instruction-following and tool-use benchmarks, the 2.6B model leads every category except one:

Benchmark	LFM2.5-2.6B	Gemma-4-E4B-it (8B)	Qwen3.5-9B (9.7B)
ToolSandbox	77.83	65.00	76.44
Multi-IF	80.07	77.35	62.55
IFStruct	85.49	76.65	78.50
IFBench	59.17	39.24	56.47
BFCLv4	56.88	46.39	60.13

Source: Liquid AI’s published benchmark results. It trails the 9.7B model only on the BFCLv4 tool-calling test, and beats it on the other four — a striking result for a model one-quarter the size. As for speed, the release reports 220 tokens per second on Apple’s M-series hardware, about 30 tokens per second on a phone, and under 2.5 GB of memory used.

When a model runs on your own device, the data-privacy conversation changes shape entirely. There is no third party holding a copy — and no way for a service provider to quietly rewrite the terms later.

Practical Takeaways

• Pick one repetitive document task — invoice data entry, order triage, or customer intake forms — and test it against LFM2.5-2.6B on a spare office PC.
• Use the post-trained LFM2.5-2.6B checkpoint for agentic tasks. Use the Base checkpoint only if you plan to fine-tune with your own data via LoRA; both TRL and Unsloth are supported.
• Tell your IT vendor or freelancer you want model files in GGUF or ONNX format, runnable in llama.cpp, vLLM, SGLang, or LM Studio. These are well-supported, non-proprietary, and don’t lock you in.
• Design workflows around the 128K context: paste entire documents in, not just extracts. That’s where the model’s real edge shows.
• Keep human oversight for anything with legal or compliance weight, especially in healthcare and financial services. The model is capable, but you still sign the report.
• Don’t use it for code generation or knowledge-heavy Q&A — that’s where larger models clearly win.

The Bigger Picture

This release is part of a broader shift: AI capability is being compressed into smaller packages, and models small enough to fit on consumer hardware are getting better at following instructions and using tools than models several times their size from a generation ago. The agentic post-training pipeline — supervised fine-tuning, per-domain teacher specialisation, distillation, and reinforcement learning inside real agent harnesses — shows that the frontier of “make AI actually do things” is increasingly a software discipline, not just a hardware arms race.

For Malaysian SME owners, the long-term implication is choice. Today, the default assumption is that good AI requires an internet connection, a foreign account, and acceptance of whatever privacy terms the provider wrote. This release demonstrates a different default: AI that lives on your own machines, speaks languages relevant to our market, and works according to the rules you set. You can decide whether that’s for you — but it’s no longer a question of whether the technology exists.

Ready to Streamline Your Operations?

Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →