Open-weight models are AI language models whose trained weights you can download, run on your own hardware and adapt to your needs. Popular free options include DeepSeek-V4, Qwen 3, GLM-5 and Llama 4, which you can run locally with tools like Ollama or vLLM, or try through cloud APIs that offer free tiers. This guide compares them and explains how to choose and run one.
Open-weight vs open-source vs closed models
- Open-weight: the model weights are published, so you can self-host and fine-tune. Training data and code may not be.
- Open-source (in the strict sense): weights plus the code and enough information to reproduce the model, under an open licence.
- Closed: available only through the provider's API.
"Free" refers to the licence: there is no fee for the weights, but you still pay for the hardware or hosting that runs them. Always read the licence before commercial use.
The popular open-weight models compared
| Model family | Developer | Licence | Known for |
|---|---|---|---|
| DeepSeek-V4 | DeepSeek | MIT | Strong reasoning and coding; large mixture-of-experts models in Pro and lighter Flash variants |
| Qwen 3 | Alibaba (Qwen team) | Apache 2.0 | The widest range of sizes, from small models that run on a laptop to large mixture-of-experts models; strong multilingual support |
| GLM-5 | Zhipu AI (Z.ai) | MIT | Coding and agent tasks with tool use |
| Llama 4 | Meta | Llama 4 Community License | Scout (109B total, 17B active) and Maverick (about 400B total, 17B active), natively multimodal; extra licence terms for services with over 700 million monthly users |
All four families are updated regularly with point releases, so check the latest version and model card before you start.
Which model should you choose?
- Run on a laptop or a single GPU: a small or mid-sized Qwen 3 model is the easiest starting point.
- Most permissive licence for a commercial product: DeepSeek-V4 and GLM-5 (MIT) or Qwen 3 (Apache 2.0).
- Coding assistants and agents: test GLM-5, DeepSeek-V4 and the larger Qwen 3 models on your own tasks.
- Images plus text: Llama 4, or vision variants in the Qwen family.
Benchmarks are a starting point only. Build a small test set from your real use case and compare two or three models on it.
How to run open-weight models locally
| Tool | Best for |
|---|---|
| Ollama | The quickest start on Mac, Windows or Linux: download and chat with one command, with a local API |
| LM Studio | A desktop app with a model browser and chat interface, no coding needed |
| llama.cpp | Efficient quantised models (GGUF) on CPUs, Apple Silicon and modest GPUs |
| vLLM / SGLang | High-throughput serving on GPU servers for production apps |
Rough hardware guide
- Small models (up to about 8B parameters), 4-bit quantised: a modern laptop with 8 to 16 GB of memory.
- Mid-size models (around 30B): a GPU with 24 GB or more, or an Apple Silicon Mac with 32 GB or more of unified memory.
- Large mixture-of-experts models such as the full DeepSeek-V4, GLM-5 or Llama 4 Maverick: multi-GPU servers or a cloud provider.
Free and low-cost cloud options
If you don't have the hardware, you can try open-weight models through hosted APIs:
- OpenRouter lists free variants of several open models alongside paid ones.
- Groq, Cloudflare Workers AI and Hugging Face Inference Providers offer free tiers or credits for selected models.
- Model developers such as DeepSeek and Z.ai run their own low-cost APIs.
Free tiers have rate limits that change often, and some free endpoints may log prompts. Use them for prototyping, not for confidential data or production traffic.
Why businesses use open-weight models
- Data privacy: keep sensitive data on your own servers.
- Cost control at scale: predictable infrastructure cost instead of per-token fees.
- Customisation: fine-tune on your own data and tone.
- No lock-in: switch models without rewriting your application.
Key takeaways
- DeepSeek-V4, Qwen 3, GLM-5 and Llama 4 are leading free open-weight families, with different licences and strengths.
- Start locally with Ollama or LM Studio, then move to vLLM or a hosted provider for production.
- Choose by testing on your own tasks, and check the licence before commercial use.
Want a private AI assistant, RAG system or agent on open-weight models? Talk to AppifyDevs.