Best open source LLMs in 2026, and what they cost to run

The short answer

If you want raw capability at data-center scale, two models lead the leaderboards in mid 2026. One is Kimi K3: Moonshot published its weights in July 2026, with a 1M token context window. The other is DeepSeek V4, and its Pro variant reports 80.6% on the SWE-bench Verified coding benchmark under an MIT license. Both of them are competitive with the closed flagships on public evaluations.

If you have a single good GPU, the Qwen mid-size dense models are the default pick. On a laptop with no GPU at all, small models like Gemma are the realistic ceiling. And if you're a developer who just wants cheap, capable inference through an API, Llama 3.3 70B and DeepSeek V3.2 answer for a tiny fraction of what the frontier models charge.

So "which open model is best" now has the same kind of answer as which AI model to use in general. It depends on the task and on where it runs. The gap to the closed models has also narrowed so far that price and license often decide it.

Open source or open weights? The license actually matters

The label "open source" covers three different deals, and you won't see the difference until it costs you.

True open licenses (MIT, Apache). DeepSeek releases its models under MIT. You can use them, modify them and build a commercial product on them, and there are no conditions worth worrying about. That's the strictest meaning of open source, and in 2026 several frontier-class models finally carry it.

Open weights with conditions. Meta's Llama license is the famous example. Almost anyone can download the weights for free, but the license restricts very large-scale commercial deployments and certain uses. The Llama explainer covers why people argue about this distinction online.

Open weights, research restrictions. Some releases allow research and personal use, but they rule out commercial deployment entirely. Read before you build.

If you only use these models through a chat product, none of this changes anything for you. If your company ships software on top of one, the license is the first fact to check, before any benchmark.

The leading open models of 2026, compared

The prices in the table are hosted API rates per million tokens, taken from the Whizi model cost index and fetched 2026-08-20 from OpenRouter list prices. The benchmark figures are the results that the developers reported on public leaderboards.

Two patterns stand out. First, 1M token context windows used to be a closed-model selling point, and now they're standard across open releases; the context window comparison tracks the full table. Second, the price floor keeps falling. Qwen3.7 Flash answers a standard query for $0.000095 on the cost index, so a thousand of its answers cost under a dime.

What do they actually cost to run?

The Whizi cost index prices every model for one standard answer, 1,000 input tokens plus 500 output tokens, so that you can compare the list rates. By that measure the open models aren't slightly cheaper than the closed flagships. They're in a different economic category.

A standard answer from Llama 3.3 70B costs $0.00026. From DeepSeek V3.2, it's $0.00047. From Kimi K3, the priciest open model in the index, it's $0.0105. The closed flagships sit at $0.0175 to $0.035 for the same answer, and across the full index the spread from the cheapest to the priciest model is roughly 2,000x. A thousand answers from Llama 3.3 cost 26 cents; the same thousand from a closed flagship can pass $30.

That's why open models quietly power the fast or economical tier of many AI products. It's also why the high-volume work almost always lands on them, whether it's classification, extraction or summarizing at scale. The live cost index has the full 100-model table with the per-answer math.

There's one caveat. Running the weights yourself is only free if your time and your hardware are. A GPU that can serve a 70B model well is a four-figure purchase, and the mid-size models that fit on one consumer GPU give up capability to get there. For most people, hosted access to open models captures the price advantage without the hardware.

Which open model for which job

High-volume, cost-sensitive work (summaries, extraction, tagging): Llama 3.3 70B or Qwen3.7 Flash. When an answer costs a fraction of a tenth of a cent, it's cheaper to be wrong now and then than to pay a premium model to be right.

Coding: DeepSeek V4 Pro or Kimi K3's coding variants. Their reported SWE-bench numbers put them in real competition with the closed coding assistants, and the MIT license makes DeepSeek the safe one to embed. Claude vs DeepSeek tests this head to head.

Long documents: any of the 1M context releases will do, and Llama 4 Maverick is the budget pick at $0.20 per million input tokens.

Local and private: a Qwen mid-size model on a 24 GB GPU, or Gemma-class small models on a laptop. Your privacy is absolute and the capability is mid-tier, and the Llama guide covers what you can realistically expect from a model you run locally.

Multilingual and European-hosted work: Mistral Large 3, which also anchors the Mistral explainer.

Where open models still lose

An honest guide names the gaps, because they're real and they decide what people buy.

The very top of reasoning. On the hardest multi-step reasoning and agentic tasks, the closed flagships still hold the edge on most published evaluations. The gap has shrunk dramatically, but it hasn't closed.

Polish and product. Open weights ship as weights. Voice modes, deep research tools, memory, and app ecosystems are all product engineering. The closed providers bundle it in, and open models leave it to whoever hosts them.

Consistency of hosting. The same open model can perform differently from one host to the next, depending on how each of them quantizes and serves it. Closed APIs don't have that failure mode. If an open model suddenly seems worse, suspect the host before you suspect the model.

Support and accountability. When a closed model misbehaves, there's a vendor to answer for it. When self-hosted weights misbehave, there's a GitHub issue tracker.

Three ways to actually use one

1. Through a multi-model workspace. This is the zero-setup route. Whizi includes open models like Llama, DeepSeek, Kimi, Qwen, and Mistral alongside GPT, Claude, and Gemini, so open-vs-closed stops being a subscription decision and becomes one you make per question. You can send the volume work to the cheap open models and the hard reasoning to a flagship in the same conversation.

2. Through an API. If you're a developer, the hosted open models at the prices above are the main reason to look past the frontier providers for your production workloads.

3. On your own hardware. Ollama and LM Studio make it genuinely easy to run a model locally now. Choose this route for privacy or offline needs, not to save money. The electricity is cheap, but the GPU isn't.