Which model for which job?
A plain-English guide to AI model types and families: large and small language models, reasoning models, open-weight models, embedding, image, speech and code models, and the main families from OpenAI, Anthropic, Google, Meta, Mistral, Alibaba (Qwen) and DeepSeek. With a job-by-job table of which kind of model to use, and a live list of the latest releases.
🎯 By the end of this lesson you will be able to:
- Recognise the main types of AI model and what each is for
- Know the major model families and their size tiers
- Match common jobs to the right kind of model, and test the choice
In this lesson
Types of AI model
| Type | What it does | Example jobs |
|---|---|---|
| Large language model (LLM) | Understands and writes text, often also images and audio | Chat assistants, drafting, summarising, analysis |
| Reasoning model / mode | Thinks step by step before answering, slower but stronger on hard problems | Complex coding, maths, planning, difficult analysis |
| Small / fast model | A cheaper, quicker LLM for simpler tasks | Classification, extraction, routing, high-volume work |
| Open-weight model | A model you can download and run yourself, under its licence | Private or offline use, custom fine-tuning, cost control at scale |
| Embedding model | Turns text into vectors for search by meaning (Lesson 8) | RAG, semantic search, recommendations, clustering |
| Reranker | Re-orders search results by relevance | Improving RAG retrieval quality |
| Image models | Generate or edit images from text | Marketing images, mock-ups, illustrations |
| Speech models | Speech to text and text to speech | Transcription, voice assistants, accessibility |
| Code models | LLMs tuned for programming | Coding assistants, code review, test generation |
The main model families
New versions arrive every few months, so it is more useful to know the families and their tiers than any single version. Most providers offer a large, most capable tier and smaller, faster, cheaper tiers.
| Family | Maker | Notes |
|---|---|---|
| GPT | OpenAI | Widely used general and reasoning models, with smaller "mini"-style variants; also open-weight gpt-oss models |
| Claude | Anthropic | Tiers from most capable to fastest: Opus, Sonnet and Haiku; strong for coding, agents and long documents |
| Gemini | Pro and Flash tiers, very long context, strong with images, video and audio; Gemma is Google's open-weight family | |
| Llama | Meta | Popular open-weight family in many sizes |
| Mistral | Mistral AI (France) | Open-weight and commercial models, including Codestral for code |
| Qwen | Alibaba | Open-weight family in many sizes, strong multilingual support |
| DeepSeek | DeepSeek | Open-weight models known for strong reasoning at low cost |
| Others | Various | Grok (xAI), Command (Cohere, enterprise RAG), Nova (Amazon), Phi (Microsoft small models) |
Which model for which job?
| Job | Kind of model to start with | Why |
|---|---|---|
| Customer chat assistant | Mid-tier LLM (e.g. Sonnet, Flash or a mid-size GPT) with RAG | Good quality at reasonable cost and speed |
| Classify or route thousands of tickets | Small, fast model (e.g. Haiku, Flash-Lite or a mini GPT) | Simple task, high volume, low cost |
| Complex coding or long agent tasks | Top-tier or reasoning model (e.g. Opus, a top GPT or Gemini Pro) | Hard problems need the strongest reasoning |
| Search your documents | An embedding model, plus a reranker | Meaning-based retrieval for RAG |
| Sensitive data that must stay in-house | Open-weight model (e.g. Llama, Qwen, Mistral, Gemma) run with Ollama or vLLM | Data never leaves your infrastructure |
| Very long documents or video | A long-context or multimodal model (e.g. Gemini Pro or Claude) | Fits large inputs in one request |
| Meeting transcription | A speech-to-text model (e.g. Whisper or a provider's speech API) | Built for accurate transcription |
| Marketing images | An image generation model | LLMs describe images; image models create them |
A simple way to choose
- Start with the job: what type of model does it need (text, embeddings, images, speech)?
- Check the constraints: data rules, budget per request, speed, languages and context length.
- Shortlist two or three models from different families or tiers.
- Test them on your evaluation set with the same prompts, comparing quality, cost and speed (Lesson 43 shows promptfoo for this).
- Pick the smallest model that passes, and re-test when new versions are released.
What this means for your role
Developer
Keep the model name in configuration, not code, so switching tiers or providers is a config change plus an evaluation run.
Tester / QA
Run the same evaluation set on every candidate model and on each new version before switching.
Business Analyst / PO
Note data sensitivity and volume for each use case; they often decide between hosted, small or open-weight models.
Tech Lead / Architect
Maintain an approved model catalogue with the job each model is approved for, and review it when major releases arrive.
Project / Delivery Manager
Smaller models can cut running costs dramatically; ask whether the top tier is really needed for each feature.
Scrum Master / Agile
When a new model is released, timebox a quick comparison on the team's evaluation set instead of switching on hype.
Try it: pick models for three jobs
- Write three AI jobs from your work, such as summarising meetings, tagging tickets and answering policy questions.
- For each, choose the model type and tier from the table, and note one constraint (cost, privacy, speed).
- Open the AI model releases page and see whether any new model might be worth testing for one of your jobs.
Key takeaways
- Model types include LLMs, reasoning models, small fast models, open-weight, embedding, reranker, image, speech and code models.
- Know the families (GPT, Claude, Gemini, Llama, Mistral, Qwen, DeepSeek) and their tiers, not just versions.
- Match the job to a model type and tier, and use the smallest model that meets your quality bar.
- Choose with your own evaluation set, and re-test when new models are released.
Key terms
- Model family
- A series of related models from one maker, released in versions and sizes.
- Tier
- A size level within a family, trading capability against speed and cost.
- Open-weight
- A model whose trained weights you can download and run yourself, under its licence.
- Multimodal
- Able to work with more than one kind of input, such as text, images, audio and video.
Check your understanding
Answer all 5 questions. Score 3 or more to complete the lesson.
1. You need to tag 200,000 support tickets into 8 categories cheaply. Which model kind should you test first?
2. Which kind of model powers search by meaning in RAG?
3. Patient data must never leave your servers. What fits?
4. What are Opus, Sonnet and Haiku?
5. What is the best final step when choosing a model?