Open Source AI Models for Solopreneurs: Cost, Licence, Fit

You already pay for Claude or ChatGPT, and every week someone says an open model would do the same work for free. Sometimes that’s true. Often it isn’t: a hosted open model costs about what the cheap closed tiers cost, the licences differ by model family (and for Hermes, by model size), and running one at home has hardware limits the hype skips. This guide prices the same job across open and closed models, lays out the licence terms from each model card, and ends with a routing rule for a one-person business.

Every price below was read on the provider’s official page on October 19, 2026, and every licence on the model’s own card. Prices change often, so treat them as “as of October 2026”.

Open source or open weight: what you actually get

Almost every “open source” model a solo business would use is really an open-weight model. You can download the trained weights, run them where you like, and build on them, under the licence printed on the model card.

That’s not the same as fully open source under the Open Source Initiative’s definition. Version 1.0 of the Open Source AI Definition asks for three things: detailed information about the training data, the complete code used to train and run the system, and the parameters. Most labs publish only the last one.

For a solo operator the distinction is mostly academic. What matters is the licence, because it tells you whether you can sell something built on the model, and what you must display when you do.

The open models worth knowing in October 2026

Seven families cover nearly every practical use. Sizes and terms come from each model card.

• gpt-oss (OpenAI). Two sizes: gpt-oss-20b, with 21B total and 3.6B active parameters, and gpt-oss-120b, with 117B total and 5.1B active. Apache 2.0.
• Qwen3.8 (Alibaba’s Qwen team). The 27B model has a native 262,144-token context and is Apache 2.0. There’s also a 2.4-trillion-parameter flagship that only a hosted provider would run for you.
• Gemma 4 (Google). Five sizes from E2B up to a 31B dense model; the small ones take 128K tokens of context and the 12B, 26B and 31B take 256K. Apache 2.0 on the card.
• DeepSeek V4.1 Flash (DeepSeek). A large mixture-of-experts model with a one-million-token context, MIT licence.
• Mistral Small 4 (Mistral AI). 119B parameters with 6.5B active per token and a 256K context. Apache 2.0.
• Llama 4 (Meta). Scout and Maverick, under the Llama 4 Community License.
• Hermes (Nous Research). Fine-tunes of other labs’ base models: Hermes 4 at 14B, 70B and 405B, and Hermes 4.3 at 36B. The licence follows the base model, which is covered below.

If you want one to try first, start with gpt-oss-20b or a small Gemma 4. Both carry Apache 2.0 terms and are small enough to run at home, so you learn what open models feel like without opening an account anywhere.

What open models cost next to Claude and GPT

The pitch for open models is that they are cheap. On a hosted API, that’s only half true.

To compare like with like, the table prices one example month of solo-business work: 300 drafting runs at about 4,000 tokens in and 1,000 tokens out each, so 1.2 million input tokens and 0.3 million output tokens. That workload is an illustration, not a measured bill. Swap in your own volume and the ratios hold.

Same job, open versus closed, list prices as of October 19, 2026 (per 1M tokens, standard rates, no batch or cache discounts):

Three things stand out.

The small closed models are as cheap as the cheap open ones. Claude Haiku 5.5 and GPT-6 Luna come in at $0.27 for the example month, a little under DeepSeek V4.1 Flash off-peak and gpt-oss-120B on Together. If cost is your only reason to switch, a smaller closed model may get you there with no new vendor.

Against the mid and top tiers, open models win clearly. The same month on Claude Sonnet 5.5 or GPT-6.1 Sol is $5.40, and on Claude Opus 5.5 it’s $10.80.

DeepSeek’s own API has peak hours. Its pricing page says off-peak rates are half the peak rates, with peak running 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. For US operators, those windows fall in the evening and overnight on the East Coast, so a nightly batch job scheduled at the wrong hour pays double.

Hermes isn’t in the table. The Hermes 4.3 card lists Nous Portal and Chutes as inference providers, but I couldn’t find a published per-token Hermes price on an official page on October 19, so it stays out rather than going in as a guess.

The honest takeaway is that an open model is a cost play only if you’re replacing a big closed model. For everything else, the real reasons are licence, control and privacy. For the closed side of the decision, the ChatGPT vs Claude vs Gemini breakdown covers the $20 subscription plans.

Licences: what you can ship

This is the part the model leaderboards skip, and the one that can bite a solo seller later. If you only use a model to draft your own emails and posts, almost any of these licences is fine. The fine print matters when you distribute something: an app with the model inside, a fine-tuned model, a paid tool built on it.

Licence check from each model card, as of October 2026:

The Hermes rows are the ones to remember. Hermes is one brand, but it’s a fine-tune of someone else’s base model, so each size inherits a different licence. Pick Hermes 4 70B for a product and you have taken on Meta’s attribution rules; pick the 14B or the 4.3 36B and you’re on Apache 2.0. The same check applies to any community fine-tune: look at the “base model” line on the card before you build on it.

The 700-million-user clause in the Llama licences will never touch a solo business. The display and naming rules can, the moment you ship.

This is a plain reading of the licence pages, not legal advice. If a product depends on it, read the full text linked from the card.

Running a model on your own computer

Local is where open models do something no closed model can: your prompts never leave the machine, there’s no per-token bill, and nobody can retire the version you rely on.

The software side is free. Ollama’s pricing page lists running models locally on its free plan and says running models on your own hardware is always unlimited; its paid Pro plan, $20 a month as of October 2026, buys credits for its cloud models, not local use. LM Studio’s free plan is described as running local LLMs on your machine, with paid tiers at $20 and $100 a month for hosted models.

Hardware is the limit. The model cards are specific:

• gpt-oss-20b runs within 16GB of memory using its MXFP4 quantization, per OpenAI’s card.
• gpt-oss-120b fits on a single 80GB data-centre GPU such as an NVIDIA H100, which is not a laptop.
• Gemma 4’s E2B and E4B have 2.3B and 4.5B effective parameters, the smallest in the family.
• Hermes 4.3 36B has quantized GGUF files in a separate repository, the format local runners load.

A five-question checklist before you install anything:

1. Is the job private enough that it shouldn’t touch a third-party server? If not, a hosted API is simpler.
2. Does the model you want fit your memory? Read the card, not a forum post.
3. Is the licence on the card one you can live with if this becomes a product?
4. Can a small model do the job? Test it on ten real examples before you depend on it.
5. Who updates it? Local models don’t change on their own, which is the point, and also means you pull new versions yourself.

If you answer “no” to the first and the fourth, stay on a hosted model. A local setup that gives worse drafts than a $0.27 hosted month isn’t a saving.

Where an open model fits next to Claude and GPT

My own stack is Claude and GPT for most of the work, Hermes alongside them, and Google Calendar holding the schedule they all feed. I’ve written up the Claude side of that operations stack separately.

The routing rule I use for a one-person operation is simple. Anything I’ll sign or that a customer reads goes to Claude or GPT, because the top closed models are still the most reliable writers and reasoners I have, and an hour lost fixing a weak draft costs more than the tokens saved. Hermes earns its seat next to them for a different reason: it’s a model I’m not renting, from a different lab, with weights that stay the same version until I decide otherwise.

That gives three working uses for an open model in a solo business:

• A second opinion from a different model family on a plan or a draft that Claude or GPT already produced.
• Bulk, repetitive jobs where a small model is good enough and the table above shows a 7 to 15 times gap to Claude Sonnet 5.5 or GPT-6.1 Sol.
• Work you want to keep off a vendor’s servers, run locally.

What I wouldn’t do is swap the main assistant for an open model to save a $20 subscription. The subscription buys a polished app, connectors and the best models; an open model buys control. Those are different purchases, and a solo operator usually needs the first one more. The lean AI tool stack shows where the paid assistant sits among the other tools.

Privacy: hosted service versus downloaded weights

“Open” says nothing about where your prompts go when you use a hosted service. That depends on who runs the server.

DeepSeek is the clearest example. Its privacy policy, last updated February 10, 2026, says it collects, processes and stores personal data in the People’s Republic of China. If customer details or anything confidential would pass through, that matters.

The fix is built into the licence. DeepSeek V4.1 Flash carries the MIT licence, so the same model is available from other hosts (Together AI lists it at $0.30 in and $1.20 out per 1M tokens as of October 2026) or on your own hardware if you have a lot of it. Read the host’s data terms either way; an open model on someone else’s server is still someone else’s server.

When to skip open models entirely

Skip them if your AI use is chat in a browser, a few dozen times a day, on a $20 plan. You would add setup and a second vendor to save nothing.

Skip them too if you need connectors to Gmail, Google Calendar or Drive working today. The closed assistants ship those; with an open model you’d be wiring them yourself.

Use one when at least one of these is true: you’re paying mid or top-tier API rates for bulk work, you need a model that won’t change under you, you’re building a product and want permissive terms, or the data shouldn’t leave your machine. More AI workflows for a one-person business are collected under AI workflows.

Pick one recurring bulk job you currently send to Claude Sonnet or GPT-6.1 Sol, run ten real examples of it through gpt-oss-120B or DeepSeek V4.1 Flash, and compare the drafts side by side. If eight of ten are usable, move that job and keep the closed model for everything you sign.