But which model suits your purpose?
Quick overview of the candidates:
Gemma 3 (Google)
- The solid all-rounder with Google DNA
- Ideal for chat, texts, summaries
- Image understanding available in 4B model and up
- Extremely long context (up to 128,000 tokens!)
- Runs efficiently quantized – even large models
- Large variants need lots of RAM/GPU
- License: open & commercially usable
DeepSeek R1 (China)
- Original model is massive, but smaller variants (e.g. 7B–14B) run exceptionally well
- Designed for reasoning: Chain-of-Thought, math, coding
- Remarkably precise and interpretable
- Base model too large – best to use distilled versions
- License: fully open (MIT)
Qwen 3 (Alibaba)
The linguistically gifted all-rounder
- From 0.5B to 32B – flexible and fast
- Supports over 100 languages
- Excellent at coding, chat and Q&A
- MoE technology enables large models with small footprint
- MoE requires special runtime tools (e.g. vLLM)
- License: Apache 2.0 – freely usable
Which model fits which use case?
If you primarily want to chat, write text, or occasionally analyze images, Gemma 3 is a strong choice. It excels with extremely long context windows and reliable Google-quality output, though the larger variants require significant computational power.
If you need an AI that excels at logical thinking, understands math, and can write code, then DeepSeek R1 is worth considering. It was specifically designed for precise reasoning and programming. While the base models are massive, there are smaller optimized versions (so-called distilled models) that also run well locally.
If you want maximum flexibility - whether for multilingual tasks, programming, or traditional chat - Qwen 3 is a true all-rounder. Thanks to its special architecture (MoE), it delivers strong performance with relatively low resource consumption. Just note that you’ll need specific tools like vLLM for the larger MoE models.
| Model | Ideal for | Strengths | Considerations |
|---|---|---|---|
| Gemma 3 | Chat, text, image analysis | Long context, Google-grade stability | Larger models require significant power |
| DeepSeek | Math, logic, coding | Strong reasoning & precision | Base model too large - use distilled versions |
| Qwen 3 | All-rounder, multilingual | Fast, flexible, multi-language | MoE variants require setup |
How to get started easily? Ollama!
Don’t want to be a terminal expert, just want to get started?
Then check out Ollama, probably the simplest tool for local AI:
What Ollama offers:
- Super simple installation (Windows, Mac, Linux)
- Automatically downloads and launches models
- Optimized for 4-bit versions (fit many GPUs)
- Many top models available: llama3, gemma, deepseek-coder, qwen, codellama, mistral, phi…
But be aware:
- No image input (even if the model supports it)
- Selection limited to pre-configured models
- Some versions may be older or only available in smaller sizes
- No fine-tuning or API like professional setups
For more advanced needs (larger models, vLLM, custom modifications), check tools like LM Studio, Text Generation WebUI or KoboldCpp.
Conclusion
Whether you want to write, code, or have smart conversations, there’s a model for your needs. Gemma 3 for solid everyday performance, DeepSeek for logic and math, Qwen 3 for languages and code. With Ollama, getting started is child’s play. Those who want to go deeper can always upgrade - locally you can run more than you might think.
