Every team building on LLMs makes this choice, often without noticing: closed models you can only reach through one vendor’s API, or open-weight models whose weights anyone can download and run. The choice affects your bill, your data path, how stable your product is, and how hard it is to leave.
What does open-weight actually mean?
An open-weight model publishes its trained parameters. You, or any provider, can run it. That is narrower than open source: the training data and full training pipeline are usually not released.
Licences vary. Some are permissive: OpenAI’s Whisper, for example, releases code and weights under the MIT licence. Others restrict commercial use, user counts or fields of use. Read the licence of each model before you build on it.
A closed model’s weights stay with the vendor. You rent access through its API, or through clouds the vendor has licensed.
Open-weight vs closed models compared
| Factor | Open-weight | Closed |
|---|---|---|
| Top-end quality | Trails the frontier by months | Usually leads on the hardest tasks |
| Price per token | Often lower; hosts compete on the same model | Set by one vendor |
| Where it runs | Your servers, a rented GPU or any host | The vendor’s API or licensed clouds |
| Privacy options | Self-host, or pick a host with zero data retention | The vendor’s terms and tiers |
| Version stability | Pin exact weights for as long as you like | The vendor retires versions on its schedule |
| Customisation | Fine-tune or quantise yourself | Only what the vendor offers |
| Lock-in | Low: several hosts serve the same weights | High: one vendor per model |
| Operations | Yours if you self-host | The vendor’s |
How big is the quality gap?
Smaller than it used to be, but real. Stanford’s 2025 AI Index reported that the lead of the top closed model over the top open-weight model on the Chatbot Arena leaderboard fell from 8.04% in early January 2024 to 1.70% by February 2025. Epoch AI estimated on May 29, 2026 that “the most capable open-weight models have lagged frontier closed models by an average of four months”, measured on its Epoch Capabilities Index.
In practice:
- Classification, extraction, summaries, translation, routine chat: open-weight models are usually good enough, and cheaper.
- Coding agents and long documents: strong open-weight options exist; test on your repos. See the best open model for coding agents.
- The hardest reasoning and newest capabilities: closed models often get there first.
Leaderboards are averages. Your 50 real test cases are the benchmark that matters.
What are the three ways to run an open-weight model?
| Option | Best when | Watch out for |
|---|---|---|
| Self-host on your own GPUs | Data must not leave your network; GPUs stay busy | Hardware, drivers, scaling, security, upgrades |
| Rent a dedicated GPU | Steady load, custom weights | You pay for idle hours |
| Hosted open-weight API | Spiky or growing traffic; fast start | You trust the host’s retention and uptime |
Self-hosting looks free per token, but a GPU costs the same whether it serves one request an hour or a thousand. It pays off at high, steady utilisation. Below that, a per-token API or a flat-rate plan is usually cheaper.
What does an open-weight model cost per month?
Per-token prices for hosted open-weight models are easy to compute. For 100 million input and 10 million output tokens a month on DeepSeek V4.1 Flash through BroadNet:
100 x $0.175 + 10 x $0.52 per month, before any cache discount.
Run the same arithmetic with your closed vendor’s published per-million prices and your real token counts, or start from our open-weight vs closed model API cost comparison, which prices one workload on GPT, Claude, Gemini, Grok and five open models. Cached input changes the result a lot: cached tokens on this model cost $0.0055 per million. The LLM API cost calculator does this for every BroadNet model, and the LLM API pricing table lists what other hosts charge for the same open weights.
Why does lock-in matter more than it looks?
When a closed vendor raises prices, changes data terms or retires the model your prompts are tuned for, your options are to accept it or rewrite. With an open-weight model, you can move the same weights to another host, or in-house, and keep the same behaviour.
To keep that option real:
- Call models through an OpenAI-compatible SDK with the base URL and model ID in config.
- Keep a regression test set and run it on every provider change.
- Optionally put a router such as LiteLLM in front, so switching is a config change.
How do businesses combine open and closed models?
Most businesses end up with both:
- Route high-volume, well-defined tasks (extraction, triage, summaries, agents’ routine steps) to an open-weight model.
- Route rare, hard tasks to a closed frontier model.
- Re-test every quarter. Given the gap Epoch measures, tasks move from the second list to the first over time.
Where BroadNet Inference fits
BroadNet Inference is a hosted open-weight API: DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B and BroadNet’s build of Google’s Gemma 4, plus two speech-to-text models, all through an OpenAI-compatible endpoint. Its privacy statement says it keeps no prompts or outputs and does not use inference data to train, fine-tune or evaluate models. DeepSeek V4.1 Flash is also available on flat plans from $1 an hour to $60 a month per session.
It does not serve closed models. If a task needs GPT, Claude or Gemini, use those vendors for that route and keep the rest on open weights. To weigh BroadNet against other providers, use the 20-point provider checklist.
