LLM Model Selection

LLM model selection is the process of choosing which large language model (a type of AI that generates text) to use for a specific task. It matters because it is a three-way tradeoff among quality (how accurate the output is), latency (how fast the model responds), and cost (how much each call to the model costs you in money and computing power).

The right choice depends on your goal. For a deterministic task (a problem with a fixed, predictable answer) that ordinary code can solve, you should not use an LLM at all—adding a model creates cost and risk without value. When you do need a model, start with the simplest shape. An augmented LLM (a model call supported by tools, retrieval, or memory) adds extra capabilities without jumping to a fully autonomous agent. A workflow is more deterministic, meaning it follows a predefined path with predictable results and gives you more control. In contrast, an agent (an LLM in a loop that calls tools until done) is more flexible but less predictable.

Stay model agnostic (able to switch between different model providers). This lets you retain ownership of your data and choose models that improve your accuracy while lowering your cost, rather than being locked into one vendor. When evaluating options, know that models tend to focus on the initial inputs and the last inputs, often removing the in-between context, so your results are only as good as the context you provide. Plan to optimize your feature later, hill climbing (iteratively improving) on an eval to tune quality, speed, and cost.

Choosing the right LLM (large language model) starts with a model picker, a dropdown menu inside tools like Claude Code or LM Studio. Your default model should handle everyday tasks, like simple code edits or answering questions. For example, set your picker to Qwen or DeepSeek for routine work—this cuts costs significantly. When something hard breaks, such as a multi-file bug, a vision task (image understanding), or a workflow heavy with MCP (tool connections), manually switch the picker to a stronger model like Opus or Sonnet. You hit the slash model command in Claude Code, then pick a model from the list.

Think of model selection as a three-way tradeoff among cost, speed, and capability. Start with the simplest model that can meet your requirement. If ordinary code solves a task deterministically (with predictable output), adding an LLM creates cost and risk without value. The model is not a mind reader—it's only as good as the context you give it, so provide detailed instructions. You can also run models locally via Ollama, which places them in the same dropdown as Claude. For coding, try Kimi K 2.6; for reasoning tasks, use the Qwen 3.5 9 billion parameter model. When evaluating, don't rely on old benchmarks—test models on your specific workload, using tools like the AI model cheat sheet bundle to understand what each model excels at. The key is to adjust your picker based on task difficulty, not to stick with one model for everything.

Choosing an LLM (large language model) is a three-way tradeoff among quality, cost, and agentic capabilities (abilities like tool calling). Quality dominates model choice for most teams, followed by cost and tool use. Open versus closed source matters to only 5% of users—don't fixate on it.

Avoid the pitfall of trusting leaderboard pass rates, which often claim 80%+ success. These numbers rarely reflect real-world performance. Instead, test models on your specific task with your data.

Be model agnostic—design your system to switch between providers like Claude, Qwen, or others. This keeps your accuracy improving and costs down. Don't marry a single vendor.

A common mistake is relying on one model to grade its own work. If you ask Claude if its output was optimal, it will almost always say yes. Use a separate judge (a different model or human review) to evaluate results.

Also, remember the model is not a mind reader—it's only as good as the context you give it. Provide clear, detailed instructions.

For local or cheaper options, Qwen 3 is a strong open-source model. Download a couple candidates, try them, and see what works best for your use case before committing.

Finally, consider if you even need an LLM. If ordinary code can solve a deterministic task (one with a predictable fixed outcome), adding a model adds cost and risk without value. Start with the simplest solution that meets your needs.