Open vs. Closed Source AI: A Decision Framework

Learning Objectives

• Explain the practical difference between open-weight and closed-source AI models
• Articulate the strategic logic behind open-sourcing vs. staying closed — and why Meta keeps its flagship proprietary while releasing smaller models under Apache 2.0
• Apply a decision framework to choose between open and closed models for specific use cases

What "Open Source" Actually Means for AI

In traditional software, "open source" means the source code is publicly available. For AI models, the term is used loosely and means different things depending on context. Not all "open" models are equally open.

The most important distinction:

Open weights: The trained model parameters (weights) are publicly downloadable. You can run the model yourself, fine-tune it, and deploy it on your own infrastructure. Examples: Llama 4, Gemma 4, DeepSeek V4, Inkling.

Open source (fully): Model weights AND training code AND training data AND detailed methodology are publicly available. This is rare. True full open source AI models (like EleutherAI's GPT-Neo series) are uncommon at frontier capability levels.

Closed source (API-only): You access the model only through the provider's API. You cannot download the weights, cannot run it yourself, cannot inspect its internals. Examples: OpenAI's GPT flagships, the Claude Opus line, Google's Gemini Pro models.

The License Spectrum

Even among open-weight models, licenses vary significantly. MIT and Apache 2.0 are the most permissive — both allow free commercial use and free modification. MIT covers Phi-4, DeepSeek V4, and DeepSeek R1; Apache 2.0 covers Mistral Small 4, Gemma 4, Qwen 3.5, GPT-OSS, Inkling, and — since August 2026 — Meta's Muse Glimmer. The Llama Community License (covering the Llama 4 series) is similar but requires a separate license from Meta once your product passes 700 million monthly active users, a ceiling almost nobody reaches. Research-only licenses block commercial use entirely and limit modification — those apply to some academic models. Closed API models — the GPT, Claude and Gemini flagships — charge per-API-call pricing and don't allow direct modification of the model.

For most commercial use cases, MIT and Apache 2.0 licenses are the most permissive and least legally complex.

Why Companies Open-Source (or Don't)

Understanding the strategic logic helps you evaluate claims about open source.

Why Meta Open-Sources

Meta's reasoning is explicit and rational — and worth reading as strategy rather than principle, because Meta's own flagship, Muse Spark, is closed while its other models are open (covered below):

1. Commoditize the complement. Meta's business is advertising on social platforms, not selling AI models. By open-sourcing AI models, Meta makes AI infrastructure cheaper and more available to everyone — which commoditizes the AI model market and prevents competitors (like OpenAI or Anthropic) from building a durable moat in foundation models.

2. Build ecosystem loyalty. When millions of developers build on Llama architecture, they learn Meta's conventions, contribute improvements back, and create a gravitational pull toward Meta's continued model development.

3. Attract and retain research talent. The open-source AI research community is large, talented, and values open publication. Meta's open approach is a talent strategy as much as a product strategy.

4. Reduce regulatory risk. Open models are harder to regulate than proprietary models — they are already distributed. Meta's open-source stance gives it a political position in AI governance debates.

Why OpenAI and Anthropic Remain Closed

1. Capability moat. Keeping model weights private prevents competitors from directly copying the most capable model. The months of lead time between a frontier model release and competitors' responses translate directly to revenue.

2. Safety concerns. Anthropic and OpenAI both argue that releasing frontier model weights creates safety risks — models can be fine-tuned to remove safety guardrails and used for harmful purposes. This argument is contested but not entirely without merit for the most capable models.

   The sharpest public rebuttal came from Mark Zuckerberg in August 2026, who argued the safety case runs the other way: that superintelligence held by "a small number of individuals, businesses, government, or AI itself" produces worse outcomes than widely distributed capability, and that treating danger as grounds for concentrating control is itself the problem. Note that this is an argument made by the company whose commercial interest is served by commoditizing models — which does not make it wrong, but is the kind of thing to weigh when a strategy is presented as a principle. Both sides of this debate are arguing from their business model as well as their convictions.

3. Commercial incentive. API access creates a clear, recurring revenue stream. Open-sourcing the model eliminates that revenue stream while still incurring the cost of developing the model.

4. Enterprise trust. Enterprise customers often prefer closed, supported models with SLAs, compliance certifications, and vendor accountability — which open weights don't provide.

When to Choose Open-Weight Models

Data privacy and sovereignty: If your application involves sensitive data — healthcare records, financial information, legal documents, personally identifiable information — and you cannot transmit that data to an external API, open-weight models run locally are the solution. You control the data; it never leaves your infrastructure.

Cost control at scale: At very high inference volumes, self-hosting an open-weight model can be significantly cheaper than paying per-token API fees. The crossover point depends on your compute infrastructure, but for millions of daily queries, the economics often favor self-hosting.

Custom fine-tuning: If your application requires specialized behavior that general-purpose models don't provide well, fine-tuning an open-weight model on your domain-specific data is often more effective and efficient than prompt engineering on a closed model.

On-device and edge deployment: Applications running on phones, embedded systems, or in environments without reliable internet access require on-device models. Only open-weight models can be deployed this way.

Research and experimentation: Academic research, red-teaming, interpretability research — all require access to the model internals or the ability to modify the model in ways that closed APIs don't permit.

Regulatory and compliance requirements: Some regulated industries and government applications require full transparency into what model is being used and full data residency control — requirements that closed APIs cannot satisfy.

When to Choose Closed-Source API Models

Maximum capability at launch: Frontier closed-source models typically lead open-weight models by 3–12 months on capability benchmarks. If you need the best available performance for a critical application, closed models usually win at any given moment.

Ease of deployment: No infrastructure to manage, no scaling challenges, no model versioning — just an API call. For most applications, this simplicity is worth the per-token cost.

Safety and alignment: Closed models have been through more extensive safety evaluation and have stronger default content policies. For customer-facing applications where harmful outputs would be reputationally damaging, the safety investment that OpenAI, Anthropic, and Google have made in their closed models is valuable.

Enterprise support: SLAs, compliance certifications, vendor accountability, dedicated support — closed API providers offer these; open-weight models don't.

Rapid iteration: When you need to build quickly and don't want to manage infrastructure, closed APIs let you focus on product, not model operations.

The Open-Source Momentum: DeepSeek and GPT-OSS Change the Calculus

For much of 2023–2024, the choice between open and closed models involved a significant capability trade-off: closed models were materially better. The landscape shifted dramatically in 2025–2026.

DeepSeek V4.1-Flash (MIT license) is DeepSeek's current model, released September 10, 2026 — a 552 billion-parameter multimodal mixture-of-experts with a one-million-token context window that activates 8 billion parameters per token while reading input. It replaced the April 2026 pair, V4-Pro and V4-Flash: V4-Flash retired the same day and V4-Pro's endpoint routes to the new model from September 14. V4-Pro carries 1.6 trillion total parameters with 49 billion active per token — since passed in raw size by Kimi K3 and Qwen3.8 Max, but still the largest open-weight model under a fully permissive license, and still downloadable even now that DeepSeek has stopped serving it. Its predecessor V3 made the argument first, matching frontier closed models at a fraction of the training cost (2.664 million H800 GPU hours); V4 is the follow-through. DeepSeek's repeated demonstration that open-weight models can match proprietary alternatives has fundamentally changed enterprise AI adoption patterns.

GPT-OSS (Apache 2.0) marks OpenAI's entry into open-weight models — a strategic acknowledgment that the open-source movement cannot be ignored even by the most commercially oriented AI lab.

Inkling (Apache 2.0) is the newest data point, and the most pointed. In July 2026, Thinking Machines Lab — founded by former OpenAI chief technology officer Mira Murati — made its first model an open-weights release rather than a closed flagship: a 975-billion-parameter mixture-of-experts system trained on 45 trillion tokens, published under Apache 2.0 with weights on Hugging Face. The lab openly concedes Inkling is not the strongest model available, open or closed, and pitches it as a base for organizations to fine-tune through Tinker, its customization platform.

That inverts the usual sequence. The established pattern is to build a closed frontier model first and open-source something smaller later, once the flagship's lead is safe. Thinking Machines skipped step one — a bet that the durable business is the customization layer rather than the base model. It is worth watching precisely because it tests a different theory of where open weights create value: not as a cheaper substitute for the frontier, but as raw material that a customer's own data turns into something the frontier cannot match on a narrow task.

The capability gap between top open and top closed models has narrowed substantially. For many use cases, the performance difference is small enough that the advantages of open deployment — privacy, cost control, customization — can dominate the decision.

Mozilla put a number on that gap in September 2026. The second edition of its State of Open Source AI report measures the lag between the open and closed frontiers at roughly 4.4 months on METR's task-horizon data, with the best open model scoring about three points behind the closed leader on the Artificial Analysis Intelligence Index at a fraction of the price. The useful part is not the headline figure but the shape underneath it, because Mozilla argues the decision is workload-specific rather than organization-specific. Closed models earn their premium in three places: tasks that would take a human expert roughly eight to twelve hours, expert professional work, and long context. Below eight hours either tier does the job, and past twelve neither does reliably. So routine work that will still be routine next quarter is the case for open weights, and a deadline that lands before the open frontier catches up is the case for paying. Mozilla also counts eight of the ten highest-volume models on the OpenRouter gateway in August 2026 as open-weight, while noting that open models still earn a small fraction of total model revenue.

Treat those as Mozilla's measurements rather than settled fact. It is an advocacy organization for open models, the report is its own reproduction of the underlying METR fit, and its headline comparison rests on benchmark scores that move every release cycle.

Meta Runs Both Tracks — Watch Which Tier Gets Which License

Meta is the best worked example in this lesson, because it does both things at once and the split is not random.

The model long rumored under the code name "Avocado" shipped on April 8, 2026 as Muse Spark, the first flagship from Meta Superintelligence Labs — and it is proprietary, built to power Meta AI across WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban glasses. Llama 4 Behemoth, Meta's largest planned open-weight model, was never released and has been deprioritized behind it. Through mid-2026 that pattern read as Meta quietly leaving the open-source movement it had done the most to build.

Then on August 10, 2026 Meta released Muse Glimmer, a 30 billion parameter model, under a plain Apache 2.0 license — no user ceiling at all, and therefore more permissive than the Llama Community License it had used for years. Mark Zuckerberg published a 6,500-word essay the same day recommitting Meta to open releases.

Two things follow. First, treating any lab's openness as permanent is a planning error in both directions — Meta tightened toward a closed flagship and then loosened to Apache 2.0 inside eighteen months, so build for the possibility that today's license is not tomorrow's. Second, the standard-bearers for open-weight frontier models remain DeepSeek, Qwen, Mistral, and newcomers like Thinking Machines. Meta is back in the open-weight conversation, but at the small-and-local end of it, not the frontier.

A Decision Framework

Ask these questions in order:

1. Can your data leave your infrastructure? If no → open weights only
2. Do you need to fine-tune on proprietary data? If yes → open weights preferred
3. Is cost at scale a primary constraint? If yes → evaluate open weights
4. Do you need maximum current capability? If yes → closed API
5. Do you need enterprise compliance features? If yes → closed API (or Bedrock/Azure OpenAI for open models with enterprise wrapping)

Most enterprise applications end up in a hybrid: closed API for complex reasoning and customer-facing intelligence, open weights for high-volume classification and extraction tasks.

Key Takeaways

• "Open source" for AI most commonly means open weights — the model can be downloaded and run locally — which is different from open training data, open training code, or open source software
• Meta open-sources Llama for strategic reasons (commoditize the complement, ecosystem building, talent), not altruism
• OpenAI and Anthropic remain closed for capability moat, safety concerns, and commercial revenue reasons
• DeepSeek V4-Pro and V4-Flash (MIT, April 2026) and GPT-OSS (Apache 2.0) demonstrate that open-weight models now compete at the frontier — V4-Pro runs 1.6 trillion total parameters with 49 billion active, the largest under a fully permissive license even after Kimi K3 and Qwen3.8 Max passed it on raw size
• Inkling (Apache 2.0, July 2026) inverts the usual order: Thinking Machines shipped open weights as its first model, betting the durable business is the customization layer rather than the base model
• Meta runs both tracks at once: the flagship Muse Spark ships proprietary, while Muse Glimmer (August 2026) is Apache 2.0 with no user ceiling — the tier has decided the license, though Zuckerberg's September 2026 promise of open weights for Muse Spark, with 1.2 planned and 1.3 undecided, means that boundary now runs between checkpoints rather than cleanly between product lines
• Check the license per model, never per company — Meta simultaneously ships closed (Muse Spark), community-licensed (Llama 4, 700 million user ceiling), and Apache 2.0 (Muse Glimmer) models
• No lab's openness is permanent — licenses follow business incentives, so plan for the possibility that today's open model has no open successor
• Choose open weights when: data can't leave your infrastructure, customization is needed, cost at scale is critical; choose closed APIs when: maximum capability, rapid deployment, enterprise compliance, or safety are priorities