“Free,” “open-weight” and “open source” mean different things
A model may be downloadable even when its training code or data is not published. “Open-weight” generally means the weights are accessible, but it does not by itself define the freedoms granted by a license. “Open source” implies broader rights to study, modify and redistribute under specific terms; in AI, a marketing label does not settle every distinction.
“Free” can also mean a hosted plan with changing limits, downloadable weights or a metered API. Even without a license fee, local inference requires hardware, power and maintenance. Before adding a model to a product, read the license and service terms for the exact checkpoint rather than relying on a family-level description.
Current releases and their specific licenses
Google announced Gemma 4 under Apache 2.0 and has also published Gemma 4 12B releases under that license. Mistral Small 4 was released under Apache 2.0; its documentation describes a multimodal model for chat, coding and reasoning. These statements apply to those releases and should be rechecked on the model page before redistribution.
The official Qwen3.8-27B model card states Apache 2.0 for that checkpoint; do not assume the same terms apply to every Qwen model. The 27B checkpoint is resource-intensive to run locally. Llama 4 uses Meta’s Llama 4 Community License, a distinct agreement that should not be described as Apache or as a conventional open-source license.
A quick view of four options
| Model | Stated license | Local use | Worth evaluating for |
|---|---|---|---|
| Gemma 4 | Apache 2.0 for the cited releases | Weights available; requirements vary by variant | Local experiments and permissive licensing |
| Mistral Small 4 | Apache 2.0 | Possible, but the full model targets substantial infrastructure | Multimodality, coding and reasoning |
| Qwen3.8-27B | Apache 2.0 for this checkpoint | Weights available; substantial local resources needed | Multilingual evaluation on suitable hardware |
| Llama 4 | Meta Llama 4 Community License | Weights available under Meta’s terms | Evaluation after reviewing the specific agreement |
This is not a ranking or hardware-compatibility guarantee. Verify the license and requirements for the exact official release.
Local does not mean cost-free or risk-free
Local inference can keep computation on a chosen device or infrastructure, but privacy depends on the whole software stack: application, telemetry, plugins and external connections. Check the data path and network settings; an interface for local models may still send requests to remote services.
Resource needs depend on architecture, precision, quantization, context length and runtime. A small quantized variant may work on compatible personal hardware, while larger models or high-throughput configurations need much more memory and acceleration. There is no universal RAM figure: check the official model card and measure the intended workload.
A checklist before commercial use
Identify the official repository and exact model ID; read the license, usage policy and terms for adapters or datasets; check attribution, redistribution, trademark and commercial-use conditions. A family name does not replace reviewing the file you actually use.
Then evaluate representative tasks, languages, quality, latency and update needs without relying only on vendor benchmarks. A permissive license does not prove accuracy, safety or suitability for sensitive data. For commercial distribution or legal questions, ask a qualified professional to review the applicable obligations.
Sources & further reading
Have a factual correction or a source to suggest? Contact the editorial desk.
