Most small businesses should not run an open-weight model on their own GPUs. If you want an open model, use it through a hosted API first, and rent GPUs only when your data must stay on hardware you control, or when you push more than about 1 to 2 billion tokens a month at a steady rate.
With our default inputs, gpt-oss-120b on one rented H100 costs $3,932 a month and starts to beat GPT-6.1 Sol at about 1.4 billion tokens a month. It never beats the same gpt-oss-120b on a hosted API at $0.15 / $0.60 per million tokens, unless the GPU is busy far more of the day. Prices checked 2026-10-14.
Who this is for: owners and product leads who keep hearing that open models are “free” and want to know if they should drop OpenAI, Anthropic, or Google. Who should look elsewhere: if you already run GPU servers and have someone on call for them, you know your own costs better than any calculator. If you just need to pick a hosted model, start with our guide to choosing an AI model.
What we have and have not done. Dawloom has not self-hosted a model for a client. Every number below comes from a vendor page, an independent benchmark, or the calculator on this page, and each one links its source.
Three ways to run a model
“Open-weight” means the company publishes the model file, so anyone can download it and run it. You still pay for the computers that run it. That gives you three options.
| Closed API | Hosted open-weight API | Rented GPUs | |
|---|---|---|---|
| Example | GPT-6.1 Sol, Claude Sonnet 5.5 | gpt-oss-120b or Kimi K3 on Together AI | gpt-oss-120b on a Lambda H100 |
| You pay for | Tokens | Tokens | Every hour the GPU is on |
| Cost when idle | $0 | $0 | Full price |
| Who keeps it running | The vendor | The host | You |
| Can you change the model weights | No | Some hosts offer fine-tuning | Yes |
| Can the model be retired under you | Yes, on the vendor’s date | Yes, the host can drop it | No, the file is yours |
The middle column is what most guides skip. Together AI, for example, sells gpt-oss-120b at $0.15 input and $0.60 output per million tokens on its pricing page (checked 2026-10-14). You get the open model and its license, and you pay only for what you use.
The break-even calculator
Enter your monthly tokens. The calculator shows the monthly bill for all three options and the volume where rented GPUs start to cost less. Every input is on screen, so you can swap in your own GPU price or measured speed. It runs in your browser.
| Option | Per month | Per 1M tokens |
|---|---|---|
| Closed API GPT-6.1 Sol | $280 | $2.80 |
| Hosted open-weight APICheapest gpt-oss-120b on Together AI | $19.50 | $0.20 |
| Rented GPUs 1 GPU at $4.29/hr, always on | $3,932 | $39.32 |
- Rented GPUs cost less than GPT-6.1 Sol from 1.4B tokens a month.
- Rented GPUs never cost less than gpt-oss-120b on Together AI at these settings.
- GPU bill: 1 GPU x $4.29 x 730 hours = $3,132, plus $800 of engineer time. These GPUs can serve about 7.1B tokens a month at 30% load.
Closed API prices from each vendor's pricing page, checked 2026-10-21. Together AI and Lambda prices checked 2026-10-14. US prices before tax. Compare models of similar quality on your own tests before you trust a cost win. It runs in your browser. Nothing you type is sent anywhere.
The same calculator lives at /tools/open-weight-break-even-calculator/. Here are its default results, for readers without JavaScript:
| Tokens a month | Closed API | Hosted open-weight | Rented GPUs |
|---|---|---|---|
| 10M | $28.00 | $1.95 | $3,932 1 GPU |
| 100M | $280 | $19.50 | $3,932 1 GPU |
| 1B | $2,800 | $195 | $3,932 1 GPU |
| 10B | $28,000 | $1,950 | $7,063 2 GPUs |
| GPU setup | GPU rent a month | vs GPT-6 Astra | vs GPT-6.1 Sol | vs GPT-6 Luna | vs same model on Together |
|---|---|---|---|---|---|
| gpt-oss-120b on 1x H100 $4.29/GPU/hr, 8,973 tokens/s per GPU | $3,132 | 281M | 1.4B | never | never 20B at 100% load |
| gpt-oss-120b on 1x B200 $6.99/GPU/hr, 44,489 tokens/s per GPU | $5,103 | 422M | 2.1B | never | 30B 30B at 100% load |
| Kimi K3 on 16x B200 $9.86/GPU/hr, 6,505 tokens/s per GPU | $115,165 | 8.3B | 41B | never | 28B 28B at 100% load |
Where each input comes from:
- GPU price. Lambda’s pricing page lists one H100 SXM at $4.29 per GPU per hour and one B200 at $6.99, on demand, US prices, checked 2026-10-14. Left on for a month (730 hours), one H100 costs $3,132.
- Speed. SemiAnalysis measures models on real GPUs in its InferenceX benchmark. It says gpt-oss-120b “sustains 8,973 tokens/s per GPU on H100 at 50 tokens/s per user,” on a chat workload of 8,000 input and 1,000 output tokens. That figure counts input and output tokens. On a B200 it measured 44,489. These are independent measurements. We did not run them.
- Load. This is the share of each hour your GPUs run flat out. We use 30%, because most business traffic comes in working hours. That is what we plan for. Change it to match your traffic.
- Engineer time. We use 8 hours a month at $100 an hour for updates, monitoring, and restarts. That is what we plan for. Set your own rate.
- API prices. Closed API prices come from the official OpenAI, Anthropic, and Google pages, checked 2026-10-13, the same file as our AI API pricing table. No caching or batch discounts, which would make the APIs cheaper still.
What the numbers say
- Load decides almost everything. At full load, one H100 serves gpt-oss-120b for about $0.13 per million tokens in rent. At 30% load, it’s about $0.44. The hosted price for the same model is $0.195 per million at a 90/10 input and output mix. So your own H100 only wins if it stays busy most of the day, and then only past roughly 20 billion tokens a month.
- Against a closed model, the GPU wins sooner. GPT-6 Astra costs $10 / $50, so one H100 pays for itself at about 281 million tokens a month. Against GPT-6.1 Sol ($2 / $10) it takes about 1.4 billion. Against GPT-6 Luna ($0.10 / $0.50) it never does at these settings.
- A cheaper model has to pass the same tests. gpt-oss-120b and GPT-6.1 Sol are different models, and gpt-oss dates from August 2025. A cost win only counts if the open model passes your own tests. Our model guide has a simple test to run first.
- Big open models need big clusters. Kimi K3 has 2.8 trillion parameters. InferenceX ran it “multi-node” on B200s. By our sizing, it needs 16 GPUs to hold it. On Lambda’s 16-GPU B200 cluster at $9.86 per GPU per hour, that’s $115,165 a month in rent. Lambda rents that cluster only for terms of 2 weeks to a year.
- Open-weight is not always cheap per token. Together charges $3 / $15 for Kimi K3, more than GPT-6.1 Sol at $2 / $10. GLM-5.3 is $1.40 / $4.40 and DeepSeek V4 Pro is $1.32 / $3.96.
Tokens are counted differently by each model, so the same text can be more tokens on one model than another. Measure your own prompts before you trust a small gap.
The license table
Open weights don’t all come with the same rights. We read each license file on Hugging Face or the vendor’s own repository on 2026-10-14. Sizes are from the model cards, except GLM-5.3, where we use Hugging Face’s parameter count. This is a summary for planning, not legal advice. If you are near any threshold, show the license to a lawyer.
| Model, current version | License | Commercial use | The catch |
|---|---|---|---|
| Kimi K3 (Moonshot AI), 2.8T total, 104B active | Kimi K3 License, custom | Yes | A “Model as a Service” business with group revenue over $20M in any 12 months needs a separate agreement. “Kimi K3” must show in your UI above 100M monthly users or $20M monthly revenue |
| GLM-5.3 (Z.AI), 753B | GLM-5.3 License, custom | Yes | Model-as-a-service firms over $10B revenue must “pass Z.AI’s security review.” GLM-5.3-Flash is plain MIT |
| DeepSeek V4 Pro 0813 and V4.1 Flash | MIT | Yes | Only keep the copyright notice. The catch is size: V4 Pro has 1.6T parameters |
| Qwen3.8 (Alibaba), three releases | Apache 2.0 (27B), Qwen Community License 1.0 (Flash-Next), Qwen3.8-Max License (2.4T) | Yes, with limits on two | Flash-Next: any model-as-a-service or “AI Work Assistant” business needs a separate license, at any revenue. Max: the same above $50M revenue. Both: model name in your UI above 100M monthly users or $20M monthly revenue |
| gpt-oss-120b and 20b (OpenAI) | Apache 2.0, plus a one-line usage policy | Yes | No license catch. Released August 2025, the oldest model here, and text only |
| Gemma 4 (Google), 31B, 26B-A4B, 12B, E4B, E2B | Apache 2.0 | Yes | No license catch on Gemma 4. Gemma 3 uses Google’s own Gemma license, so check which version you deploy |
| Llama 4 Scout and Maverick (Meta) | Llama 4 Community License | Yes, with conditions | Show “Built with Llama.” Fine-tuned models must start with “Llama.” Over 700M monthly users needs Meta’s permission. EU companies get no rights to the multimodal models. Downloads are gated. Released April 2025 |
| MiMo-V2.6-Pro (Xiaomi), 1.02T total, 42B active | MIT, per the model card | Yes | The repository has no LICENSE file. The only license statement we found is the “license: mit” tag on the card. Not confirmed beyond that |
Every model in the table is the newest version its maker has on Hugging Face as of 2026-10-14. Meta has published nothing newer than Llama 4, and OpenAI nothing newer than gpt-oss. New open models ship every few weeks, so check the dates.
The license lines that matter
Most of these licenses are short. These are the lines a business should read in full.
Kimi K3. The service clause reads: “If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars … the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.” Two exceptions matter for most businesses. A product with AI “solely embedded within specific features” is not a model service. And “internal use,” meaning nothing reaches third parties, is exempt. Wikipedia says the license requires “revenue sharing of up to 30%” for inference providers earning over $20 million a year. The license file has no such line, so we treat the 30% as not confirmed.
Qwen3.8. The trap is the middle model. The Qwen Community License says that if you run a model service or an “AI Work Assistant” business, “the licensee shall obtain a separate license from Qwen before Using the Software.” It has no revenue floor. An AI Work Assistant is “an independent AI-powered product primarily designed for AI-assisted coding or office productivity.” If you are building an AI coding or office tool, use the Apache 2.0 Qwen3.8-27B, or ask Qwen first.
Llama 4. The license says that if you share a product built on it, you must “prominently display ‘Built with Llama’.” The acceptable use policy adds: “the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union.” Scout and Maverick are both multimodal, so an EU company can’t use them under this license. Its end users can still use a product built on them.
GLM-5.3. The security review applies only above $10 billion of revenue, so it won’t touch a small business. The license also says the scope of the review “shall be reasonably determined by Z.AI.” If you need a GLM model with no extra terms, GLM-5.3-Flash and GLM-5.2 are MIT.
gpt-oss, Gemma 4, DeepSeek V4. Plain Apache 2.0 or MIT. OpenAI’s model card calls it a “Permissive Apache 2.0 license: Build freely without copyleft restrictions or patent risk.” Its usage policy asks only that you “comply with all applicable law.”
Good reasons to skip the API
Cost alone rarely justifies your own GPUs below a billion tokens a month. These reasons can:
- Data must stay on your hardware. Some contracts or regulators require that no outside company processes the data. Rented GPUs or your own servers solve that. A hosted open-weight API does not.
- No internet. Air-gapped sites, ships, and factories need the model on site.
- You need to change the model. Fine-tuning on your own data, or a model the hosts don’t offer.
- You can’t accept a forced change. Closed models retire on the vendor’s date. Our model retirement calendar lists the ones coming up. A model file you keep never changes.
- Steady, very high volume. Batch jobs that run around the clock keep the GPU busy, which is when the math works.
Weak reasons: “open models are free” (the GPUs aren’t), and “we’ll save money at our current volume” (check the calculator first). Before any of this, ask whether the feature needs a model at all. Our checklist for AI features helps with that.
What other guides leave out
We read the top results for “self-host LLM vs API” before writing this. The three we read in full, from Cloudzy, Digital Applied, and White Beard Strategies, give break-even numbers but none states the GPU speed behind them. None quotes a license beyond noting that Qwen3.8-27B is Apache 2.0, and none mentions the Qwen Community License or the Llama 4 EU clause. One sells GPU servers and two sell consulting. We sell AI work too, so weigh that. This page shows every input, links each price, and quotes each license.
If you are working out what AI will cost per customer, our AI cost per user guide takes the next step. If the model sits behind an MCP server, where to host an MCP server covers the server side.
FAQ
Can I use Kimi K3 commercially?
Yes. The Kimi K3 License allows commercial use, changes, and resale. Two limits apply. A model-as-a-service business with group revenue over $20 million in 12 months needs a separate agreement with Moonshot AI. Products with over 100 million monthly users or $20 million monthly revenue must show “Kimi K3” in the UI. Internal use is exempt from both.
Is self-hosting an LLM cheaper than using the API?
Usually not. With our default inputs, one H100 running gpt-oss-120b costs $3,932 a month including engineer time. It beats GPT-6.1 Sol only from about 1.4 billion tokens a month, and never beats the same model on a hosted API at 30% load. Use the calculator with your own numbers.
What GPU do I need to run gpt-oss-120b?
One 80 GB GPU. OpenAI’s model card says gpt-oss-120b fits “into a single 80GB GPU (like NVIDIA H100 or AMD MI300X)” and gpt-oss-20b runs “within 16GB of memory.” Lambda rents one H100 for $4.29 an hour, checked 2026-10-14.
Which open-weight models have no commercial catch?
Of the eight families we checked, gpt-oss, Gemma 4, DeepSeek V4, Qwen3.8-27B, and GLM-5.3-Flash use Apache 2.0 or MIT with no extra terms. MiMo-V2.6 is tagged MIT but ships no license file. Kimi K3, GLM-5.3, the larger Qwen3.8 models, and Llama 4 add conditions.
Is an open-weight model safer for my data than ChatGPT?
Only if you run it on hardware you control. A hosted open-weight API still sends your data to that host, so read its data terms the same way you would OpenAI’s or Anthropic’s.
If you want a second opinion on which model and setup fit your product, that is part of our AI integration work. We test candidate models on your own data before you commit to any of them.