The Orchestration / Tool-Calling article covered how a model takes real, structured actions. This article shifts to a genuinely different, ecosystem-level question — who controls the model itself — starting with the side of that question defined by openness: open-source models.
The simple definition
An open-source model makes its trained weights, and often its full training code, freely available for anyone to download, run, modify, and redistribute. Recall from the Model Poisoning article’s discussion of platforms like Hugging Face, back in the Responsible AI phase — open-weight models are exactly the kind of downloadable files that entire ecosystem is built around. Once downloaded, an open model belongs, in a real, practical sense, to whoever downloaded it — no ongoing dependency on the original creator’s servers staying online, or their pricing staying the same.
Why access to the actual weights changes what’s genuinely possible to build
Recall from the CUDA article’s discussion of ecosystem advantage, and the Quantization and INT4 articles, back in the Infrastructure & Serving phase — llama.cpp’s entire premise depended on having direct access to Llama’s actual weight files, something no closed-source model permits. An open-weight model can be fine-tuned on private, sensitive data without that data ever leaving a company’s own servers, quantized down for a specific device, or run entirely offline with no internet connection at all — genuine capabilities a closed, API-only model structurally cannot offer, regardless of how good its terms of service happen to be.
flowchart LR
A[Open-Source Model] --> B[Weights downloaded directly]
B --> C[Fine-tune privately, quantize, run fully offline]
B --> D[No ongoing dependency on the original provider]ANALOGY vs. TECHNICAL REALITY
Analogy: Think of buying a car you can open the hood of, modify, and repair yourself, rather than leasing one serviced exclusively by the manufacturer. Once it’s yours, it stays yours, regardless of what the manufacturer decides to do next.
Where this breaks down: A car’s hood is either literally open or literally locked. As the real controversy below demonstrates, “open” for an AI model exists on a genuinely contested spectrum, with real, serious disagreement about where the actual line should be drawn.
The real, ongoing fight over what “open source” actually means for AI
This deserves telling in full, because it’s a genuine, unresolved controversy involving real organizations and real financial stakes. Meta has marketed its Llama models as open source since Llama 2’s 2023 release.
But Llama’s actual license carries real restrictions the traditional definition of open source has never permitted: it forbids using Llama to train competing models, and requires companies with more than 700 million monthly active users to obtain a special license before using it at all — provisions specifically written to block Meta’s largest competitors.
Erica Brescia, a managing director at RedPoint, asked publicly: “Can someone please explain to me how Meta and Microsoft can justify calling Llama 2 open source if it doesn’t actually use an OSI-approved license?” The Open Source Initiative — the nonprofit that has formally stewarded the definition of “open source” for over 25 years — published a blog post titled, without hedging, “Meta’s Llama 2 license is not open source.” In 2024, OSI released a formal definition of open-source AI specifically requiring that training data itself be genuinely accessible, a standard Llama has never met.
The real, 2025 case that showed what open weights can actually do
This is worth telling as a real, dramatic counter-example, because it demonstrates what downloadable weights can enable. DeepSeek, a Chinese AI startup, released R1 in January 2025 with downloadable weights under a permissive license. It is more accurate to call R1 open-weight than to claim that every part of its creation was fully open.
You may see a figure of roughly $5.6 million connected to DeepSeek. That figure described DeepSeek-V3’s reported final pretraining run; it was not the complete cost of researching, staffing, collecting data for, testing, and creating R1.
The market’s reaction was immediate and historic: Nvidia’s stock fell nearly 17% in a single day on January 27, 2025, erasing roughly $593 billion in market value, the largest single-day market-cap loss for any company in U.S. stock market history.
DeepSeek followed with a genuinely aggressive real pricing strategy — publicly listing API rates around 0.55 per million input tokens, with an additional off-peak discount of up to 75% introduced weeks later, compared to 15 per million tokens charged by several major closed-source competitors at the time.
Because its weights were downloadable, independent researchers worldwide could run, modify, and benchmark it instead of depending only on DeepSeek’s API. That access does not automatically reveal every training example, experiment, or total development expense.
A concrete example, layered
For a simple beginner example: a hospital building a patient-intake assistant that must never send data to an external server would likely choose an open-weight model specifically so the entire system can run on the hospital’s own private infrastructure. For a production example: DeepSeek’s real, published models and Mistral’s Mixtral, referenced throughout the Generative Models phase, represent genuine, more permissively usable alternatives many developers have specifically chosen over Llama, partly in direct response to the licensing controversy covered above.
Why open weights still don’t mean full, verified transparency
It’s worth being honest about a genuine, remaining gap here. A 2023 study co-authored by researchers at Carnegie Mellon, the AI Now Institute, and the Signal Foundation found that many models branded “open source” share a real limitation — the training data stays secret, and the compute needed to meaningfully retrain or audit the model remains far beyond most developers’ reach, meaning “open” in practice often means “downloadable and runnable,” not the fuller transparency the term has traditionally implied in software.
The real cost of self-hosting
Downloading weights removes the provider’s per-token bill, but it does not make operation free. A team must still pay for GPUs, electricity, storage, networking, monitoring, security, engineers, and spare capacity for busy periods.
downloaded weights
↓
GPU servers + serving software + engineers + monitoring
↓
monthly self-hosting costFor a small and uneven workload, a hosted API may cost less because the team pays only when it sends requests. Self-hosting becomes attractive when privacy, offline operation, customization, or a large predictable workload matters enough to justify the infrastructure.
Separate the pieces before calling a model open
| Piece | What it lets you do |
|---|---|
| Weights | Run the already-trained model. |
| Inference code | Turn inputs into outputs using those weights. |
| Training code | Repeat or adapt the training process. |
| Training data information | Study what the model learned from and reproduce it more closely. |
| License | Know what modification, redistribution, and commercial use are legally allowed. |
Before deployment, record the exact repository revision and verify the downloaded file’s cryptographic hash. This is like checking a sealed package number: it helps prove that every server received the intended file rather than an altered or accidentally different copy.
Common misconception
Open source and open weight are not always the same
Open weight means people can obtain the trained parameter files. Open source AI may additionally require usable code, documentation, license permissions, and enough information to study or reproduce how the system was made.
weights available? -> can the model run locally?
license permissions? -> can it be modified or used commercially?
training information? -> can people understand how it was produced?
source code available? -> can the software be inspected and changed?Do not decide from the word “open” in a marketing page. Read the actual license and model card.
A practical download checklist
- Confirm the publisher and exact repository revision.
- Read commercial-use, redistribution, and acceptable-use terms.
- Inspect the model card, intended uses, known limitations, and evaluations.
- Prefer safer weight formats and isolate unfamiliar loading code.
- Evaluate quality, bias, security, latency, and hardware cost yourself.
Llama, Gemma, DeepSeek, and many Qwen releases provide downloadable weights, but they do not all use identical licenses or disclose identical training details. Hugging Face lists license metadata separately because “downloadable” does not answer what a user is legally allowed to do.
Verified sources
A frequent beginner assumption: that a model labeled “open source” by its own creator has met some formal, verified, industry-standard bar for openness. As the OSI-versus-Meta dispute directly demonstrated, this isn’t a safe assumption — it’s genuinely worth checking a model’s specific license terms rather than trusting the label alone.
Where this fits in what comes next
You now understand what genuine openness offers, and how contested the label itself has become. The next article, Closed-Source Models, covers the other side of this same spectrum — and a real, direct admission from OpenAI about exactly why it stopped disclosing what its own models are actually built from.
In one sentence
Open models can expose different combinations of weights, code, data information, and permissions; DeepSeek R1 showed the practical power of permissively licensed downloadable weights, while the continuing Llama debate shows why the actual license and released artifacts matter more than the marketing label.
Related Terms
- Author
- TechByteByByte Editorial Team
- Reviewed by
- TechByteByByte Admin
- Published
- Last reviewed