...choose a model
Jump to navigation
Jump to search
| AI Solver map |
|---|
|
Your current page is in bold
0 · Start
1 · Predict & analyze
tables, numbers, events
Predict values
Predict categories Forecast time series Cluster / segment Find outliers Make recommendations Find key drivers 2 · Understand unstructured data
text, documents, images, audio
3 · Generate & converse
4 · Decide & act
5 · Build it right
applies to every path
|
YouTube ... Quora ...Google search ...Google News ...Bing News
- AI Solver ... Model Search ... Discriminative vs. Generative ... ...customize a model ... ...deploy AI ... ...understand new architectures
Use this page when you know your goal and need to pick a model, especially a language, vision or speech model. New models appear every few weeks, so this page teaches a method that stays valid, plus the current landscape.
Start here: the checklist
- Is it really a language-model task? Numbers in a table → classic ML usually wins (...predict values, ...predict categories). Use LLMs for language, documents, images, audio, code and reasoning over them.
- Modality → text only, or images, audio and video in and out? Choose a multimodal model, or a specialist for one modality (...understand images and video, ...work with speech and audio).
- Difficulty
- Simple and high-volume (classify, extract, route, rewrite) → small, fast tier ("mini", "flash", "haiku"-class models) or a small language model (SLM)
- Multi-step reasoning, math, hard code, planning → a reasoning ("thinking") model; set the reasoning effort to match the task
- Raw speed matters more than top quality → speed-optimized models, including diffusion LLMs
- Data control
- Data may go to a vendor under a no-training, short-retention contract → API models (OpenAI, Anthropic, Google, xAI, Mistral, Cohere, Amazon Nova), directly or through a cloud platform (Amazon Bedrock, Microsoft Foundry, Google Gemini Enterprise Agent Platform)
- Data must stay on your own infrastructure or offline → open-weight models you host (families such as Qwen, DeepSeek, Kimi, GLM, Llama, Gemma, Mistral, gpt-oss, NVIDIA Nemotron, IBM Granite, Microsoft Phi). Check the license: Apache 2.0 and MIT are permissive; some families have custom terms
- Context size → very long inputs: long-context models, or retrieval (...answer questions from my data)
- Tools and structure → agents need reliable tool calling and JSON output; check the Berkeley Function-Calling Leaderboard and test on your tools
- Where it runs → phone, laptop or edge: SLMs of roughly 1–15 billion parameters, quantized (...deploy AI)
- Cost and latency → compare cost per completed task (including retries, reasoning tokens and tool calls), not just price per million tokens
The method (works with any model, any year)
- Write an eval set first: 30–100 real examples with the expected result → ...evaluate and trust AI
- Prototype with the most capable model to learn the quality ceiling.
- Try cheaper, smaller models and keep them wherever they still pass.
- Route: send easy requests to a small model and escalate hard ones to a large model.
- Re-test when new models ship, and keep your code model-neutral so switching is a configuration change.
The old rule of thumb still holds, refined: use small models for high-precision, low-latency, narrow or structured steps (classification, extraction, tool-argument formatting); use large models for broad reasoning, cross-domain knowledge and multi-step planning; hybrid systems route between them.
Model tiers (October 2026)
| Tier | Good for | Example families (check leaderboards for current versions) |
|---|---|---|
| Frontier, closed | Hardest reasoning, coding, agents, multimodal | OpenAI GPT, Anthropic Claude, Google Gemini, xAI Grok |
| Frontier, open-weight | Self-hosting, customization, data control | DeepSeek, Qwen, Kimi, GLM, Llama, Mistral, gpt-oss |
| Small / edge (SLM) | Low latency, on-device, narrow tasks, agent sub-steps | Gemma, Phi, small Qwen and Llama models, Granite, SmolLM, Apple's on-device model |
| Specialized | One job done well and cheaply | Embedding and reranking models, speech-to-text and text-to-speech, vision detectors, time-series and tabular foundation models |
Use leaderboards to shortlist, not to decide
- LMArena — human preference votes (text, vision, web development)
- Artificial Analysis — quality, speed and price compared
- SWE-bench — coding agents fixing real GitHub issues
- Berkeley Function-Calling Leaderboard — tool use
- MTEB — embedding models
- Open ASR Leaderboard — speech recognition
- OpenRouter rankings — which models developers actually use
- Hugging Face models — sort by trending or downloads
Real-world examples
- Routing saves money — LMSYS's open-source RouteLLM learned when a cheaper model is good enough, and reported large cost reductions while keeping about 95% of GPT-4's quality on standard benchmarks. RouteLLM | LMSYS
- Small on the device, large in the cloud — Apple Intelligence runs a roughly 3-billion-parameter model on the device and a larger model on Apple's servers; app developers can call the on-device model offline, with no per-request cost, through Apple's Foundation Models framework. Foundation Models framework | Apple Developer
Watch out
- Benchmarks leak and saturate — public test sets end up in training data; trust your own evals.
- Lock-in — avoid features that only one vendor has unless they are worth it; keep prompts and evals portable.
- Model updates — hosted models change; pin versions where possible and re-run evals on updates.
Learn more
- A practical guide to building agents (model selection section) | OpenAI
- Choosing a generative AI service | AWS Decision Guides
- Model Search ... Algorithms ... Discriminative vs. Generative
Where next? ◀ Build it right · ...customize a model · ...deploy AI · ...understand new architectures · ▶ ...evaluate and trust AI