Last week Meta released Llama 2, and for the first time a genuinely capable large language model came with a license that lets most companies use it commercially, for free. That is a bigger deal than the usual model-release noise, and it is worth understanding even if you never touch the thing yourself.

For most Indonesian SMEs, the honest answer is that a good open source llm for business use will not change what you do next month. You should probably keep using an API like the ones from the major providers. But the arrival of Llama 2 changes the market underneath you, and it opens a real door for a specific kind of business. Both of those are worth ten minutes of your attention.

Let me strip the romance out of this, because there is a lot of it going around.

What Llama 2 actually is, in plain terms

Llama 2 is a language model whose weights, the actual trained brain of the thing, you can download and run on your own hardware. Compare that to the well-known hosted models, where you send your text to someone else's servers and get an answer back. With a hosted API you rent intelligence. With an open model you own a copy.

The comparison people jump to is: is Llama 2 as good as the top hosted models? For the largest hosted models, not quite, not yet. For a lot of ordinary business tasks, it is close enough that the gap does not matter. Summarizing documents, drafting replies, classifying support tickets, extracting fields from messy text. On those, an open model does the job.

The word "free" is doing a lot of work, though, and it is the most misunderstood part.

Free to license is not free to run

Downloading Llama 2 costs nothing. Running it in production costs plenty, and it is not the kind of cost that shows up as a tidy monthly bill.

Here is what self-hosting actually means for a business:

  • Hardware. Capable models want serious GPUs. Buying them is a large capital cost; renting them from a cloud provider adds up fast and can easily exceed what you would have paid an API for the same volume.
  • Engineering. Someone has to deploy the model, keep it running, handle scaling, patch it, and fix it at 2am when it falls over. That is a real salary or a real retainer, not a hobby.
  • Ongoing maintenance. Models improve, security issues appear, dependencies break. Self-hosting is a system you now own forever, with all the responsibility that implies.

An API, by contrast, is a per-request cost with none of that operational weight. You pay for what you use and someone else keeps the lights on. For most SMEs doing modest volumes, the API is not just easier, it is genuinely cheaper once you count the hidden costs honestly. This is the same trap as any tempting shortcut, and it is closely related to why cheap software ends up costing more: the sticker price hides the real bill.

Who genuinely needs an open model

So who should actually care about self-hosting? A specific group, and if you are not in it, you can move on with a clear conscience.

  1. Data-sensitive industries. If you handle medical records, financial data, or anything under strict confidentiality rules, sending customer text to a third-party API may be a compliance problem or simply unacceptable to your clients. Running the model inside your own walls solves that cleanly. A multifinance company handling sensitive borrower data is a classic example.
  2. High, steady volume. At small volumes, APIs win on cost. At very large, predictable volumes, the math can flip and owning the infrastructure becomes cheaper. You have to actually do the calculation, not assume.
  3. Deep customization needs. If you need to fine-tune a model heavily on your own domain and keep full control of that tuned version, owning the weights matters. Most businesses do not need this. Some genuinely do.

If none of those describe you, the API is the right call, and choosing it is not settling. It is good engineering discipline.

Why it matters even if you never self-host

Here is the part that affects everyone, including the API crowd.

A capable free model is downward pressure on the entire market's prices. When anyone can run a decent model themselves, the hosted providers cannot charge whatever they like. Competition from open models has already pushed API prices down and will keep doing so. You benefit from Llama 2 existing even if you never download it, because it makes the service you do use cheaper and forces the whole field to move faster.

It also changes negotiating leverage. If a vendor tries to lock you into their proprietary model at a premium, "we could self-host an open alternative" is now a credible position, even if you would rather not. Options you do not exercise still shift the deal.

The practical takeaway

Llama 2 makes a strong open source llm for business use freely licensable, and that quietly improves your position whether you adopt it or not. But do not let anyone romance you into self-hosting. For most SMEs, an API remains simpler and cheaper once the hidden costs of hardware, engineering, and maintenance are counted honestly.

Self-host only if you handle genuinely sensitive data, run very high steady volume, or need deep customization. Otherwise, use the API, enjoy the falling prices that open models are driving, and put your energy into the actual business problem the AI is meant to solve. If you want help running that decision for your specific situation instead of guessing, that is the kind of call I help clients make as a technology partner.