Open Weights vs Closed Models — Choosing Where AI Lives

Status: 🟩 COMPLETE 🟦 LIVING Tags: open-weights, open-source, closed-source, proprietary, llama, gpt, self-hosting, privacy


What it is

When people talk about AI models being “open” or “closed,” they’re describing who can access the model itself — the underlying mathematical weights that encode everything the model has learned.

  • Open-weights models: The trained model is released publicly. Anyone can download the weights, run the model on their own hardware, modify it, fine-tune it, or build products with it. You can use it without sending data to anyone.

  • Closed models / proprietary models: The model is kept private. You can only access it by sending your data to the company’s API (via the internet), paying for access, and trusting them with your data. Examples: GPT-4o (OpenAI), Claude (Anthropic), Gemini (Google).

This distinction has profound implications for privacy, cost, control, and capability.


A concrete analogy

Closed model = visiting a restaurant. You go to their location, they cook with their kitchen and their recipes (which are secret), you eat, you pay, they keep the recipe.

Open-weights model = getting the recipe. You receive the actual recipe (the model weights). You can cook it at home, modify it to your taste, serve it to others, or build a business around it. The restaurant no longer controls what you do.


Major open-weights models (mid-2026)

ModelCreatorCountryBest for
Llama 3.3 70B / Llama 4Meta🇺🇸General purpose; most popular
Mistral 7B / Mixtral 8x7B / Mistral Large 2Mistral🇫🇷Quality + efficiency; European
Gemma 2 / Gemma 3Google🇺🇸Small, efficient; on-device
Phi-3 / Phi-4Microsoft🇺🇸Tiny but capable; mobile
Qwen 2.5Alibaba🇨🇳 ⛔Good quality — but Chinese; avoid
DeepSeek R2DeepSeek🇨🇳 ⛔Good quality — but Chinese; avoid
Command R+Cohere🇨🇦RAG-optimised; enterprise
Falcon 2TII🇦🇪Research use; UAE

Major closed models (mid-2026)

ModelCreatorCountryAccess
GPT-4o / o3 / o4OpenAI🇺🇸ChatGPT; OpenAI API
Claude 3.5 / Claude 4Anthropic🇺🇸Claude.ai; Anthropic API
Gemini 2.5 Pro / FlashGoogle🇺🇸Gemini; Google API
Grok-3 / Grok-4xAI🇺🇸X Premium; xAI API
Mistral Large 2Mistral🇫🇷Mistral API (also available via cloud)

The key differences

DimensionOpen-weightsClosed
PrivacyMaximum — data never leaves your machineData sent to provider’s servers
CostFree (you pay compute)API usage fees
CapabilityCatching up; approaching frontierCurrently top performance
CustomisationFine-tune, modify anythingLimited to fine-tuning on provider
SpeedDepends on your hardwareGenerally fast (optimised cloud)
ReliabilityYour infrastructure’s reliabilityProvider SLA
Vendor lock-inNone — you own the weightsDependent on provider
UpdatesYou choose when to updateProvider updates (may change behaviour)
Chinese modelsAvailable but risky (DeepSeek, Qwen)Assessed per vendor

When open-weights is clearly better

1. Maximum privacy requirements

You’re processing medical records, legal documents, financial data, or anything deeply sensitive. The information cannot leave your infrastructure under any circumstances.

Open-weights means: the AI runs on your servers, your laptop, or your cloud account. Data never touches the model company’s servers.

2. Air-gapped environments

Government, defence, or critical infrastructure environments with no internet access. The only AI you can use is the one you’ve downloaded.

3. Fine-tuning for a specific domain

You want to train the model further on your company’s documents, customer conversations, or specialised knowledge. Open-weights gives you full control over the training process.

4. Cost at very high volume

For very high-volume applications (millions of daily queries), running your own open-weights model can be significantly cheaper than paying API fees, if you have the technical expertise.

5. Regulatory requirements

Some regulations (healthcare, finance, EU GDPR in certain interpretations) may require demonstrating that data doesn’t leave specific jurisdictions or infrastructure. Self-hosted open-weights satisfies this.


When closed is clearly better

1. Capability at the frontier

For the most demanding tasks — advanced reasoning, complex creative work, cutting-edge multimodal tasks — frontier closed models (GPT-4o, Claude 3.5, Gemini Ultra) currently outperform available open-weights models, though the gap is closing.

2. Zero technical setup

Closed models via APIs are available immediately, with no infrastructure to set up. Open-weights requires servers, GPUs (or a strong CPU), and technical knowledge to deploy.

3. Small usage volume

If you’re sending a few hundred queries per day, API costs are minimal and self-hosting is overkill.

4. Specialised capabilities

Some closed models have unique features (DALL·E 3 image generation tightly integrated with GPT, Claude Artifacts, Gemini’s multimodal capabilities) that don’t have direct open equivalents.


Running open-weights models locally: the tooling

The main tools for running open-weights models locally (no GPU required for smaller models):

ToolWhat it doesDifficulty
OllamaEasiest way to run models locally; one command⭐ Beginner
LM StudioGUI for running models locally; very accessible⭐ Beginner
GPT4AllDesktop app; simple; private⭐ Beginner
vLLMProduction-grade server; high throughput⭐⭐⭐ Advanced
llama.cppEfficient inference; broad hardware support⭐⭐ Intermediate

For Australian users with a modern laptop (8GB+ RAM): Ollama + Llama 3.2 3B or Phi-3 Mini can run locally with acceptable speed for testing.


The “open-weights” vs “open-source” distinction

These terms are often confused:

  • Open-source (true): Full access to training code, training data, and model weights. Anyone can reproduce and verify everything. Very rare in practice for frontier models.

  • Open-weights: Only the trained model weights are released. The training code and data may or may not be released. You can run and modify the model but can’t reproduce the training.

Most models called “open” are open-weights, not truly open-source. Meta’s Llama models, for example, release weights but not training data. This matters because:

  • You can’t independently verify what the model was trained on
  • You can’t retrain from scratch
  • There may be biases or backdoors in the training you can’t audit

Australian government and regulatory angle

The Australian Government’s AI frameworks increasingly consider data sovereignty:

  • The Digital Transformation Agency’s cloud strategy encourages keeping sensitive data in Australian or Five Eyes-nation infrastructure
  • Open-weights models deployed on Australian cloud infrastructure (AWS Sydney, Azure Australia East) satisfy many data residency requirements
  • Chinese open-weights models (DeepSeek, Qwen) should be treated with the same caution as Chinese closed models — see vendors-chinese-avoid

Gotchas

  • “Open” doesn’t mean safe by default. Open-weights models can contain biases, produce harmful content, and require the same careful deployment practices as closed models.
  • Local models require hardware. Running a 70B parameter model (Llama 3.3 70B) requires roughly 40GB of GPU VRAM or specialised CPU memory. Not every computer can do this.
  • Smaller local models are less capable. A Phi-3 Mini (3.8B parameters) running on your laptop is significantly less capable than GPT-4o. Don’t expect frontier-model quality from laptop-size models.
  • Fine-tuning is technical. Customising an open-weights model requires significant ML engineering expertise — not a no-code experience.
  • Open weights ≠ no risk. You’re still responsible for AI outputs. Running your own open-weights model doesn’t eliminate hallucinations, bias, or harmful output risks.

See also


Sources

  • Touvron et al., “Llama 2: Open Foundation and Fine-Tuned Chat Models” (Meta, 2023)
  • Jiang et al., “Mistral 7B” (2023)
  • OSI (Open Source Initiative) definition of open source vs open weights
  • Australian Digital Transformation Agency — cloud and data sovereignty policy
  • Bommasani et al., “On the Opportunities and Risks of Foundation Models” (Stanford, 2021)
  • IEEE Spectrum — “The Open vs Closed AI Debate” (2024)