Open-Weight vs Closed AI Models: Which Should a Startup Build On?
Choose open-weight models when control, cost and data handling outweigh the convenience of a managed API, and closed models when you want the provider to carry operations and much of the compliance work. Most startups should test both on their own tasks and keep a swap path, because the gap is now narrow enough that neither choice is permanent.
Key takeaways
- Open-weight means the trained weights are published; open-source usually adds code, data and licence terms; closed models are reached only through a provider.
- Kimi K3.1 Pro shows how close the gap is: a vendor-reported Terminal-Bench 4.0 score of 65.7% against 66.4% for Claude Opus 5.5, with full weights promised by 4 Nov 2026.
- Price no longer separates the camps: Kimi lists $2 input and $10 output per million tokens, the same as gpt-6.1-sol and half of Claude Opus 5.5 according to Simon Willison.
- Self-hosting adds operations, evaluation and security work that a closed API hides.
- Procurement risk cuts both ways: a US DHS order bars federal civilian agencies from third-party products built on DeepSeek technology.
Choose open-weight models when control, cost and data handling outweigh the convenience of a managed API, and closed models when you want the provider to carry operations and compliance work. Most startups should test both on their own tasks and keep a swap path, because neither choice is permanent.
TL;DR
- Open-weight means the trained weights are published; open-source usually adds code, data and licence terms; closed models are reached only through a provider.
- Kimi K3.1 Pro shows how close the gap is: a vendor-reported Terminal-Bench 4.0 score of 65.7% against 66.4% for Claude Opus 5.5, with full weights promised by 4 Nov 2026 (Kimi blog).
- Price no longer separates the camps: Kimi lists $2 input and $10 output per million tokens, the same as gpt-6.1-sol and half of Claude Opus 5.5 according to Simon Willison (Kimi pricing, Simon Willison).
- Self-hosting adds operations, evaluation and security work that a closed API hides.
Disclosure: ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 26 Oct 2026; ZeroRaised has no commercial relationship with the companies named.
Open-weight models are models whose trained parameters are published so that anyone can download and run them, while closed models are reached only through a provider's API and the weights stay with the vendor.
What do open-weight, open-source and closed mean?#
Open-weight is the narrower term: the trained parameters are released, so you can run the model on your own hardware or through a host. Open-source usually means the weights plus the code, data and licence terms needed to reproduce or retrain the model. The labels get used interchangeably in marketing, so check what is published.
Closed models sit at the other end. You call an API and the provider runs the model. You do not get the weights, you cannot self-host, and your data passes through their systems under their terms.
| Term | What is published | Who runs it | What you control |
|---|---|---|---|
| Open-weight | Trained weights | You or a host | Hosting, data path, some tuning |
| Open-source | Often weights plus code, data and licence | You or a host | The full stack, subject to the licence |
| Closed | API access only | The provider | Prompt, integration and model choice |
How close are open-weight models to the frontier?#
Kimi K3.1 Pro, released by Moonshot AI on 23 and 24 Oct 2026, is the current test case. Moonshot calls it "our most capable 3T-class open-weight model to date, with a 1-million-token context window" (Kimi blog). On its vendor-reported Terminal-Bench 4.0 score it reaches 65.7%, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra (Kimi blog). Treat vendor benchmarks as claims, not verdicts.
The naming needs care. At publication the weights were not yet downloadable: Moonshot said they would be released by 4 Nov 2026, and Simon Willison noted on 23 Oct that they had not been released (Kimi blog, Simon Willison). Until then, K3.1 Pro is a closed API with an open-weight promise.
Moonshot says its hallucination figures come from an internal stress test, where claim-level hallucination fell from 23.4% on K3 to 8.3%, not everyday error rates (Kimi blog).
What does each option cost?#
| Model (per million tokens, Kimi pricing, 24 Oct 2026) | Input | Output | Cached input |
|---|---|---|---|
| kimi-k3.1-pro | $2.00 | $10.00 | $0.20 |
| kimi-k3.1-pro Highspeed | $4.00 | $20.00 | $0.20 |
Moonshot lists flat per-token pricing without context-length tiers, a 1,048,576-token context and Zero Data Retention for API customers (Kimi pricing, Kimi blog).
Closed pricing is competitive. OpenAI lists gpt-6.1-sol at $2.00 input and $10.00 output per million tokens with cached input at $0.10, and gpt-6-luna at $0.10 and $0.50; batch and flex tiers are half price (OpenAI API pricing). Simon Willison put the Kimi price at "the same as GPT-6.1 Sol and half the price of Claude Opus 5.5" (Simon Willison).
What are the trade-offs?#
Cost and margin. Closed APIs are predictable per token and leave the hardware to the provider. Open weights can be cheaper at steady high volume, but the saving depends on utilisation: an idle GPU is expensive. OpenAI applies a 10% uplift for regional processing endpoints on models released on or after 5 Mar 2026 (OpenAI API pricing).
Control and portability. Closed models give you a prompt, an API and the provider's parameters. Open weights let you pin a version, run it offline, fine-tune it and move hosts, which matters when a model is deprecated or limits change.
Privacy and data handling. With a closed API, your data leaves your infrastructure under the provider's terms; Moonshot says Zero Data Retention is available for its API customers (Kimi blog). With open weights you can keep data in your own environment, but self-hosting moves security work to you, including access control, logging, patching and key management.
Operations and skills. A closed API is one integration. An open-weight deployment is a service: capacity planning, evaluation after updates, latency tuning and on-call. For a solo founder, that is often the deciding factor.
Procurement and jurisdiction. The DeepSeek order shows why model choice is now a procurement question. A Department of Homeland Security order requires federal civilian agencies to remove DeepSeek-related technology from IT systems and bars procuring DeepSeek products and services, including third-party products developed using its technology (CNBC). The Federal Register notice covers "third-party products and services developed using DeepSeek implementations" under the Federal Acquisition Supply Chain Security Act of 2018 (Federal Register).
It was issued on 24 July and made public on 19 Oct, applies to civilian agencies such as Treasury, State and Justice, and excludes the Department of Defense. DHS cited potential data transfer to China and security vulnerabilities (CNBC). If you sell to government or regulated buyers, you need a model bill of materials and a swap plan.
A practical decision checklist#
- Classify each workload, separating low-risk, high-volume tasks from sensitive ones.
- Run your own evaluation, then calculate cost per task including retries and tool calls.
- Check data terms, retention and key handling (see our API key rotation guide).
- Check procurement exposure, including models inside third-party tools.
- Keep a swap path behind a small interface and revisit when the weights ship.
Where ZeroRaised fits#
ZeroRaised is a launch site for products built without institutional venture capital or accelerator equity. If your product is built and you have made a deliberate model choice, submit it. The pricing page explains the listing options, and the newsletter carries new analysis and launches. Nothing here promises traffic or revenue.
Model choice also affects marketing. How to market an API covers explaining technical trade-offs to buyers, and how to price an AI product covers turning model costs into a price that holds. For the cost side, how to reduce LLM costs goes deeper on caching and batch tiers.
Sources#
Frequently asked questions
What is the difference between open-weight and open-source AI?
Open-weight means the trained parameters are published so you can download and run the model. Open-source is often used more broadly to mean the weights plus the code, data and licence terms needed to understand or reproduce the model. Check what is actually published before you rely on either label.
Is open-weight cheaper than closed AI?
Not automatically. The token rate can be lower or equal, and self-hosting can be cheaper at steady high volume. But you also pay for compute, engineering time, evaluation, monitoring and incident response. Compare cost per completed task, not the price per token.
Can I self-host an open-weight model?
Yes, if the weights are actually available and you have the hardware and skills to serve them. That means capacity planning, evaluation after updates, latency tuning and on-call. For a solo founder, the operations burden is often the deciding factor.
Is open-weight safer for private data?
It can be, because you can keep data inside your own environment. That also moves security work to you. With a closed API, the provider processes your data under its terms, and some providers offer stronger options such as zero data retention.
Should a bootstrapped startup use open-weight models?
Test them, and avoid committing the product to one model. Use your own evaluation set, calculate cost per task and check procurement exposure. An open-weight model is a real option when the weights are downloadable, a host can serve them and the quality holds on your workload.