Topics

On this page

Open Source vs Closed Source LLMs

Open source vs closed source LLM is the choice between language models whose trained weights are published for anyone to download and models that are accessible only through a provider's hosted API. The choice determines where inference runs, who controls the data, how costs scale and how far the model can be customised.

Open-weight models: Their parameters can be downloaded and run on private hardware, for example Llama, Mistral and Gemma models.
Closed models: Their weights stay with the provider, and applications send requests over an API, for example Claude, GPT and Gemini.
Weights: They are the learned numbers inside a large language model that determine its behaviour.
Licence terms: Open-weight licences range from permissive licences to custom licences with usage restrictions.
Hosting responsibility: Self-hosted models require the team to provision GPUs, serving software and monitoring.
Terminology: Most "open source" LLMs are more precisely open-weight, because training data and training code are often unpublished.

Where inference runs for a closed source LLM and an open source LLM

For example, a team adding pull request summaries to an internal code review tool can call a closed model's API or run an open-weight model on its own GPU server.

Quick Answer

Closed source LLMs are usually faster to adopt and offer strong general capability without infrastructure work. Open source LLMs, more accurately open-weight models, give full control over data location, customisation and long-term cost, but the team operates the infrastructure. Individual model families are compared in popular LLMs compared.

Open Source vs Closed Source LLM: Comparison Table

Aspect	Open source (open-weight) LLM	Closed source LLM
Access to weights	Downloadable	Not available; API access only
Where inference runs	Own servers, private cloud or a laptop	Provider's infrastructure
Data handling	Prompts can stay inside the network	Prompts are sent to the provider under its data terms
Cost model	Hardware and operations cost, independent of request volume	Usage-based billing, typically per token
Customisation	Full fine-tuning and modification possible	Limited to prompts and any tuning options the provider offers
Setup effort	High: serving, scaling, security and upgrades	Low: an API key and a client library
Model updates	Team decides when to upgrade	Provider updates or retires model versions on its own schedule
Licence	Varies by model; some restrict commercial use	Governed by the provider's terms of service

When to Use Open Source LLMs

Strict data residency: Source code or customer data must not leave the organisation's network.
Heavy, predictable volume: Large constant workloads can cost less on owned hardware than on per-token billing.
Deep customisation: The task requires fine-tuning on internal code, domain vocabulary or a fixed output style.
Offline or edge use: The model must run without internet access, often as one of the small language models.
Version stability: The application needs a model that never changes unless the team upgrades it.

When to Use Closed Source LLMs

Fast prototyping: A working feature is required quickly, without GPU procurement or serving expertise.
Top general capability: The task demands advanced reasoning or long-document analysis where frontier hosted models are commonly preferred.
Variable traffic: Usage is irregular, so paying per request avoids idle hardware costs.
Managed features: The team needs built-in tool calling, safety filters and monitoring from the provider.
Small teams: No engineers are available to operate model infrastructure.

Example: Pull Request Summaries Both Ways

The same feature, summarising a pull request diff for reviewers, is built with each type of model.

With a closed source LLM:

Integration: The review tool sends the diff and an instruction to the provider's API and receives a summary.
Data path: The diff leaves the company network, so the provider's retention and training terms must be reviewed.
Cost: Each summary is billed by input and output tokens, so large diffs cost more.
Maintenance: The provider handles scaling and upgrades, but a model version can be retired with notice.

With an open source LLM:

Integration: The team downloads an open-weight model and serves it with an inference server such as vLLM or Ollama.
Data path: The diff never leaves the internal network, which satisfies strict code-confidentiality rules.
Cost: Cost depends on GPU hardware and engineering time rather than on the number of summaries.
Maintenance: The team monitors latency, applies security patches and chooses when to adopt newer models.

Key Differences in Practice

Capability gap: The difference between the strongest open and closed models changes frequently, so current evaluations on the team's own tasks matter more than general reputation.
Hybrid designs: Many systems route sensitive or simple requests to a self-hosted model and complex requests to a hosted API.
Hallucination risk: Both types can produce confident errors, as explained in LLM hallucination.

Quick Quiz

Pick an answer to check yourself. Nothing is saved.

Question 1 / 3

1. What does an open-weight LLM make available?
2. Which requirement most strongly favours an open source LLM?
3. How is a closed source LLM usually billed?

Frequently Asked Questions

Open weight vs open source: what is the difference?

An open-weight model publishes its trained parameters, while a fully open source model also publishes training code and usually the training data. Most well-known open LLMs share weights but not their full training data.

Are open source LLMs cheaper than closed source LLMs?

It depends on volume. Self-hosting has fixed hardware and staffing costs, so it can be cheaper for heavy, constant workloads and more expensive for small or irregular ones.

Can open source LLMs be used commercially?

Many can, but each model has its own licence. Some licences are permissive, while others add conditions such as usage restrictions, so the licence must be read before deployment.

Which is more secure, an open source or a closed source LLM?

A self-hosted LLM keeps data inside the network, which reduces exposure to third parties. The team then carries full responsibility for securing the servers, access and updates.