LLM fundamentals
Frontier vs open-weight
Frontier models arrive through a vendor API. Open-weight models are downloadable parameters, not automatically open source.
Two ways to get a strong model: call a vendor, or run weights you can download. The quality ceiling, the ops burden and the legal terms are different. Most products start on an API and stay there until spend or compliance forces a second look.
Concept
Weights are the learned numbers inside the neural network, billions of floats that define the model. "Open weights" means you can download those numbers. It does not automatically mean you get the training data, the full training recipe, or legal freedom to use the model for any commercial purpose. Always read the license.
Frontier closed models
Frontier usually means the strongest models available only through a vendor API: Claude (Anthropic), GPT (OpenAI), Gemini (Google). You send tokens, get tokens back, and never see the weights. You get frequent upgrades, tool-use and multimodal features first, and managed reliability. You pay per token (or a subscription wrapper), accept the vendor's data policies, and live with model IDs that deprecate.
These models dominate when quality and speed-to-ship matter more than unit cost or data residency: customer-facing chat, hard reasoning, coding agents that must hit high tool-call accuracy.
Open-weight families
Open-weight means the vendor publishes the trained parameters, often under a license that is not OSI open source. Llama (Meta), Qwen (Alibaba), DeepSeek, Mistral, Gemma (Google) are the families you will see most. You can run them on your GPUs, fine-tune them, or rent them from a host (Together, Fireworks, Groq, Bedrock, Azure and others).
Open weights win when you need on-prem or VPC isolation, custom fine-tunes, predictable capacity, or aggressive cost at high volume after you have proven the task. They lose when you need the absolute best quality tomorrow morning with no ops team.
Tiers inside a family
Every serious provider ships a ladder: large / medium / small. Names rotate: Opus/Sonnet/Haiku, GPT-class vs mini, Gemini Pro vs Flash, Llama 70B vs 8B.
| Tier | Role | Typical use |
|---|---|---|
| Large / frontier | Highest quality, highest $/token and latency | Hard reasoning, ambiguous prompts, agent planning |
| Mid | Best default for most product features | Chat, RAG answers, tool calling in production |
| Small / flash | Cheap and fast | Classification, routing, extraction, rewriting |
API vs self-hosting
| Managed API | Self-host / dedicated | |
|---|---|---|
| Time to first feature | Hours | Days to weeks |
| Peak quality access | Frontier models | Open-weight ceiling, or a private fine-tune |
| Unit economics at huge volume | Can get expensive | Can win after GPU plus engineering cost |
| Ops burden | Low | High. See cost per token |
| Data control | Contractual | Physical |
Trade-off
Default to a managed API until monthly spend and compliance pressure justify a serving team. Many teams stay on APIs forever and still hit their margin targets with caching, routing and smaller models. Self-host when you have stable high QPS, a model that fits your quality bar, and people who can run GPUs.
Common mistakes
- Treating "open weights" as open source. Read commercial-use clauses, attribution and cloud-provider restrictions.
- Self-hosting in week one. You have not proven the task yet.
- Ignoring residency and retention. Region pinning, zero-retention and training-on-your-data defaults belong in the same review as quality.
Check yourself
- You can download the parameters but not the training data, and the license forbids your SaaS use case. Open-weight, open source, or both? Open-weight, not open source. The license still binds you.
- Quality is unproven and you have no GPU team. API or self-host? API. Prove the product first.
Key takeaways
- Frontier APIs hide the weights and maximize speed-to-ship.
- Open-weight means downloadable parameters, not a blank-check license.
- Use the mid tier as the default. Reserve frontier for the hard slice.
- Default to an API until volume or compliance forces serving.
Further learning
Next, models that spend extra tokens thinking: reasoning models. For GPU economics, see cost per token.
Related
The LLM landscape and model choice
Frontier vs open-weight models, reasoning models, and a process for picking one with your own evals.
Choosing a model
Prove the product on a strong model, freeze an eval, then step down the ladder. Your suite decides, not a leaderboard.
The economics of inference
Turn $/GPU-hour into $/million tokens with utilization and throughput, then decide API vs self-host and price products with real margins.