Notice: This article was created with AI.
What’s It About?
The world of artificial intelligence is no longer limited to OpenAI’s GPT models. A growing number of alternative large language models (LLMs) is significantly expanding the range of available AI systems. They come from technology corporations such as Meta, Google and Alibaba, from specialised companies such as Anthropic and Mistral, and from providers like DeepSeek and Moonshot. The decisive difference today lies less in the underlying architecture than in three practical questions: are the weights openly available or not? What does the licence permit? And does the model run in the provider’s cloud or on your own hardware?
Background & Context
Updated August 2026. The original version of this article described the model landscape of 2023 – Alpaca, Vicuna, Orca and Falcon. Those were the interesting alternatives at the time, but they no longer play a role. This section has therefore been rewritten from scratch.
Among the closed models, three providers are the serious alternatives to OpenAI. Anthropic offers the Claude family with a flagship (Opus), a balanced standard option (Sonnet) and a variant geared towards long contexts and writing. Google covers the range with Gemini, from the very fast Flash model to the Pro variant with search integration. xAI positions Grok through its direct link to X and through markedly looser content filters.
The bigger shift of the past two years, though, is in open weights. Meta’s Llama is far from alone there: Alibaba’s Qwen, DeepSeek from China, Mistral from France, Google’s Gemma and Kimi from Moonshot all ship models whose weights you can download and run on your own hardware. In coding and step-by-step reasoning, these open models now sit close to the leading closed ones.
What matters here is the licence, not the word “open”. The range runs from genuine Apache 2.0 and MIT – as with Qwen and DeepSeek – to bespoke community licences with conditions attached, as with Llama and Gemma. For private experiments that makes no difference; for commercial use it very much does.
The way these models are built has shifted too. The strongest open models today are mostly mixture-of-experts systems: they hold a very large number of parameters in total but activate only a small share of them per request. That cuts the compute cost per answer considerably – but it still demands a lot of memory, because the entire model has to be held in place. A model with hundreds of billions of parameters will therefore not run on a laptop, even if only a fraction of it does the computing.
At the other end are compact models in the range of a few billion parameters. They do not reach the peak scores of the large systems, but for well-defined tasks they are often enough – and the data never leaves the building.
What Does This Mean?
- Open weights are no longer a fallback. If data protection, cost or sovereignty means you do not want to sit in a US provider’s cloud, you no longer automatically give up quality.
- The licence decides, not the label. “Open” does not mean the same thing everywhere. Read the licence before any commercial use – Apache 2.0 and MIT are unproblematic, community licences can carry conditions.
- Large does not have to mean expensive to run, but it always means expensive in memory. Mixture-of-experts lowers the cost per answer, not the hardware requirement. If you want to self-host, budget for the whole model, not for the active parameters.
- For many tasks a small model is enough. A compact model with well-prepared data of your own regularly beats a large model without that context.
- The market moves faster than any overview. Between the first version of this article and this update, the models named back then were replaced entirely. Specific version numbers have a short shelf life – the selection criteria do not.
Sources
- Large Language Models: 12 GPT-Alternativen (Computerwoche)
- 14 LLMs that aren’t ChatGPT (InfoWorld)
- Beyond ChatGPT: A Guide to Alternative Large Language Models (Teneo.ai)
- ChatGPT Alternatives (V7 Labs)
This article was created with AI assistance and is based on the cited sources as well as the language model’s training data.
Further Reading: GPTs, Skills, Plugins, Agents – Who Offers What, and What’s Actually Worth It?