Open models in October 2026: free to download, but is running them yourself cheaper than an API?

On 15 October, Meta released Muse Spark Open — an 816-billion-parameter model anyone can download. A few weeks earlier, Xiaomi released MiMo-V2.6-Pro, the open model at the top of the independent rankings. The news usually comes with the word "free", and many executives naturally ask: why not run it ourselves, stop paying for an API, and stop sending data outside?

A fair question. The answer depends on three numbers the news rarely mentions: how much GPU memory the model needs, what renting that many GPUs costs per month, and how many tokens the business actually processes per month. This article works out all three.

Four open models worth knowing

• MiMo-V2.6-Pro (Xiaomi, 21 September). About 1 trillion parameters, using only about 42 billion per step (a mixture-of-experts architecture — many sub-"experts", only a few called each time). MIT licence. Top of the open models on Artificial Analysis's index at about 46 points. API about $0.43/$0.87 per million input/output tokens.
• DeepSeek V4.1 Flash (DeepSeek, 10 September). About 552 billion parameters, 1-million-token context. DeepSeek's API, per the price schedule cited by several sources: $0.15/$0.60 off-peak, double at peak times.
• Muse Spark Open (Meta, 15 October). 816 billion parameters, about 42 billion active. Meta reports about 39 points on the Artificial Analysis index itself; no independent measurement existed when this was written. Meta's own licence, with conditions — see below.
• Qwen3.8-27B (Alibaba, August). 27 billion parameters, "dense" (every parameter is used each step). Apache 2.0 licence. Small enough to run on a single 24 GB GPU when compressed to 4-bit.

For context: the leading closed model on the same index, Claude Opus 5.5, scores 58. The best open model is about 12 points behind — a much narrower gap than a year ago, but still a gap.

"Open" does not mean unconditional

These models are often called "open source", but they are more precisely open-weight models: you can download the trained parameters, but you do not necessarily get the training data or unconditional rights. The conditions live in the licence, and licences differ a lot:

• MIT and Apache 2.0 (MiMo, Qwen3.8-27B): commercial use, modification and bundling into products almost without restriction.
• Meta's Muse licence: according to analyses of the licence, products must display "Built with Muse Spark Open"; a company offering the model as a hosted service to others with more than $20 million revenue over 12 months needs a separate agreement with Meta (internal use or building it into your own product is not subject to that threshold); the acceptable-use policy prohibits the unauthorised practice of regulated professions such as finance, law and medicine; it is not intended for users under 18; and downloading requires your full name, date of birth and organisation name.

None of these conditions is unusual. But a clinic planning to use a model to help read lab results, or a financial advisory firm, needs to read the regulated-professions clause carefully before choosing. Read the licence before you read the leaderboard.

Calculation one: how many GPUs

Most GPU memory goes to holding the model's weights. A rough rule: number of parameters × bytes per parameter.

• At native BF16 precision: 2 bytes per parameter.
• Compressed to 4-bit: about 0.5 bytes per parameter, at a slight cost in quality.

Applied:

• Qwen3.8-27B: about 54 GB at BF16 (fits one 80 GB GPU), about 16 GB at 4-bit (fits one 24 GB GPU).
• MiMo-V2.6-Pro, 1 trillion parameters: about 500 GB even at 4-bit.
• Muse Spark Open: the smallest official checkpoint is about 530 GB; reports cite a reference setup of eight H200 GPUs (141 GB each) or four new-generation Blackwell GPUs.

Add memory for context (the KV cache) — with long contexts and many simultaneous users, that can reach tens to hundreds of GB. First conclusion: the leading open models need a data-centre GPU cluster, not a server in the back office. A 27-billion-parameter model is different — that is something a mid-sized business can run itself.

Calculation two: monthly GPU rent

Cloud GPU rental prices vary widely by provider. According to Mercatus's September 2026 price index, compiled from dozens of providers, the average price per GPU per hour is roughly: H100 $3.84 (ranging from $1.79 to $12.52), H200 $4.43, B200 $6.39. A month of continuous running is about 730 hours.

• 1 × H100 (enough for Qwen3.8-27B at BF16): about $2,800 a month.
• 8 × H200 (enough for a model the size of Muse Spark Open or MiMo-V2.6-Pro): about $25,900 a month.

That is just the hardware. It excludes the engineers who install, monitor, update and handle incidents at midnight — for an eight-GPU cluster running a real service, that is usually at least part of an experienced engineer's job.

Calculation three: break-even in tokens

Now compare with calling the same model through an API. MiMo-V2.6-Pro via API costs about $0.43/$0.87. Assuming 3 input tokens for every output token (common when a model reads more than it writes), the blended price is about $0.54 per million tokens.

Break-even = monthly GPU rent ÷ API price per token:

• $25,900 ÷ $0.54 per million ≈ 48 billion tokens a month.
• Equivalent to about 18,000 tokens per second, continuously, 24 hours a day for 30 days.

For comparison, the shop with 30,000 messages a month in the article on small models processes about 72 million tokens a month. Through the MiMo API, that is about $39. Running the same model yourself costs about $25,900 — roughly 660 times as much.

Small models tell the same story. One H100 running Qwen3.8-27B costs $2,800 a month; against a small model via an API at $0.10/$0.50 (about $0.20 per million tokens blended), break-even is about 14 billion tokens a month — still nearly 200 times the shop's volume.

One more factor tilts the calculation further: rented GPUs bill by the hour whether or not anyone is using them. An internal system busy only during 8 office hours has a real cost per token three times the figures above. An API only charges when there is a request.

A detail specific to businesses in Vietnam

DeepSeek's API price schedule, as cited by several sources, has peak hours from 01:00 to 04:00 and from 06:00 to 10:00 UTC on weekdays, at double the price. Converted to Vietnam time (UTC+7), that is 8:00–11:00 and 13:00–17:00 — almost exactly office hours.

So a Vietnamese business using DeepSeek for daytime work will mostly pay the peak price, not the off-peak figure usually quoted in the news. Conversely, batch jobs (summarising documents, classifying an old data archive) can be scheduled for evenings or nights to get the off-peak rate. Check the hours on the official pricing page before scheduling.

When self-hosting is right

The money rarely favours self-hosting at the volumes of small and medium businesses. But there are legitimate reasons that have nothing to do with money:

• Data must not leave your infrastructure. Law, customer contracts or industry rules may forbid sending certain data to third-party servers. Then self-hosting — or renting dedicated infrastructure in a permitted location — may be the only option, and the cost is a compliance cost.
• No internet, or fixed latency required. Production lines, field equipment, systems that must keep running when international cables are cut.
• Deep fine-tuning needed. Further training on your specialist data, beyond what vendors' APIs allow.
• Control over versions and price. Closed models can be retired or repriced; an open model you have downloaded stays where it is.

Between the two ends are middle options: calling the open model through a provider's API with servers in a region that fits your data requirements; renting a dedicated server that runs the model only for you; or self-hosting a small model like Qwen3.8-27B for sensitive data and using an API for the rest.

Five questions before deciding

One — does the licence allow what you plan to do? Read it in full, especially the clauses on fields of use, revenue and attribution.

Two — how many GPUs does the model need? Parameters × 2 bytes (BF16) or × 0.5 bytes (4-bit), plus context. Check against the reference setup on the model page.

Three — what is the monthly rent? Number of GPUs × hourly rate × 730. Get real quotes from two or three providers; prices vary several times over.

Four — what is your real volume in tokens per month? Take it from your current API invoices or estimate from requests × average length. Compare with the break-even point.

Five — if the maths does not favour it, is there another reason? Write it down specifically: which law, which contract, which operational requirement. If you cannot write it down, use the API and revisit in six months.

Conclusion

The open models of October 2026 are genuinely good news: the gap with the leading closed models has narrowed, licences like MIT and Apache allow free commercial use, and the option to self-host makes the whole market more competitive.

But "free to download" does not mean "free to run". For the leading open models, the hardware to run them yourself would pay for tens of billions of API tokens a month — hundreds of times what most small and medium businesses need. The right reasons to self-host usually lie in data, operations or control — not in the invoice.

If you need to assess whether to deploy an AI model on your own infrastructure — from licensing and GPU configuration to a hybrid of self-hosting and API — contact us or see KonexForge's AI & ML capability.