Understand supports several local large language models (LLMs) for code explanation, navigation help, and exploratory analysis. The “best” model depends on your goals, hardware, and any compliance requirements you may be working under.
This article is here to help you choose deliberately, not guess.
Note: visit our Hugging Face page to find direct links to the models we use!
Quick Recommendation (If You Just Want an Answer)
Most users: Stick with the default Qwen3-1.7B
Better reasoning, stronger answers: Qwen3-4B or Qwen3-8B
Strict US-origin software requirements: Llama-3.2-3B or Llama-3.2-8B
Low-memory systems: Qwen3-0.6B
High-end machines / deeper analysis: Qwen3-14B
If you want full control, you can also:
Load your own GGUF model
Connect Understand to Ollama or LM Studio
Why Qwen Is the Default
In our internal testing and real-world usage, Qwen models consistently perform better than Llama models for code understanding tasks, especially:
Explaining unfamiliar code
Following multi-step instructions
Staying grounded in the codebase instead of hallucinating
That’s why as of build 7.2.1242 Understand defaults to Qwen3-1.7B — it’s the best balance of quality, speed, and size for most users.
That said, defaults are just starting points.
Model Options Explained
Qwen Models (Default & Recommended)
Best for: General use, code comprehension, accuracy, instruction following
Qwen3-0.6B (378 MB)
Fast and lightweight
Useful for quick summaries or low-memory systems
Limited reasoning depth
Qwen3-1.7B (1.0 GB) ← Default
Strong baseline for most workflows
Good explanations without heavy hardware demands
Qwen3-4B (2.3 GB)
Noticeably better reasoning and context retention
A great upgrade if you have the RAM
Qwen3-8B (4.7 GB)
Much stronger at cross-file reasoning
Slower, but more “thoughtful”
Qwen3-14B (9 GB)
Best Qwen option for deep analysis
Requires a capable machine
Overkill for simple queries, excellent for complex systems
Llama Models (US-Origin Alternative)
Best for: Compliance-driven environments, US-only software requirements
Some organizations require models developed by US companies. In those cases, Llama is the right choice, even if it’s not always the strongest performer for code tasks.
Llama-3.2-1B (1.3 GB)
Comparable in size to Qwen3-1.7B
Less consistent in instruction following
Llama-3.2-3B (3.2 GB)
Best overall Llama balance
Recommended Llama starting point
Llama-3.2-8B (5.3 GB)
Stronger reasoning than smaller Llama models
Still generally behind Qwen at the same size
Why choose Llama anyway?
Regulatory or contractual requirements
Internal policy on model provenance
Consistency with other tooling already using Llama
Choosing Based on Your Needs
Choose a Smaller Model If:
You’re on limited RAM
You want fast, lightweight answers
You mostly ask simple questions or summaries
Choose a Larger Model If:
You work in large or complex codebases
You want better cross-file reasoning
You rely heavily on AI explanations during exploration
Choose Llama If:
You must use US-origin software
Compliance matters more than raw output quality
Using Your Own Model or External Runtimes
Understand isn’t limited to the bundled models.
You can also:
Load your own GGUF file
Useful if you’ve standardized on a specific model internally.Connect to Ollama or LM Studio
Ideal if you already run models centrally or want easy model switching.
This flexibility lets you align Understand with your existing AI stack instead of reshaping your workflow around it.
Final Thoughts
There’s no single “best” model — only the best fit for how you work and where you work.
If you want reliability and strong code understanding, start with Qwen.
If you need compliance and provenance guarantees, Llama is there for you.
And if you’re experimenting or optimizing, Understand gives you room to explore.
If you ever feel like the AI is almost right but not quite — that’s usually a signal to change model size, not abandon the feature.
And if you want help tuning that choice, you know where to find us.