The model landscape

In the LLM call we dialled a number: model. So far that field just held a name, and someone at the other end picked up. Now let's look at who you can actually call - because the choice is far bigger than the famous names suggest.

Two worlds: API and open weights

The first fork isn't "which model" but "where does it run". API models (Claude, GPT, Gemini) live with their provider: you send your request over, pay per token, and get the strongest model of the day without ever touching a graphics card. In exchange, your data leaves the house - and the provider can change prices, behaviour or the model itself at any time.

Open-weights models (Llama, Qwen, Mistral, Hermes …) can be downloaded - the weights, the actual numbers of the network, end up as a file on your disk. They run on your hardware, your data stays with you, and the model never changes without you. In exchange, you are now the operator.

By the way: "open weights" is not the same as "open source". You get the finished numbers, but rarely the training data or the recipe they were made with.

Size: what "7B" means

The B stands for billions - parameters, the learned numbers in the network. A 7B model has seven billion of them, a 70B model ten times as many. More parameters roughly means more world knowledge and better reasoning - but also more memory, more compute, more cost per answer.

Still, the reflex "take the biggest" is often wrong. For "summarise this email" or "is this answer a yes or a no?", a small model is faster, cheaper - and practically just as good. The skill is knowing your task, not the leaderboard.

Families and variants

Models come in families: same architecture, different sizes (Llama 3.1 ships as 8B, 70B and 405B). Within a family you'll meet suffixes:

• Base - the raw model straight out of training. Completes text, but doesn't hold conversations.
• Instruct / chat - post-trained to follow instructions. This is what you want for agents.
• Specialists - for code, for images (vision), for embedding text (embeddings).

So which one do I take?

All of this clicks fastest when you see it on real models. That's what the panel below is for: not a ranking but a spectrum - from a 3B model for your notebook to the flagship that only runs in a data centre.

Browse

Twelve open models to look at

No ranking and no substitute for benchmarks - a shop window. Each tile shows this chapter's terms on a real model; one click opens the write-up and the way to the repo.

As of 2026-07-25 - every repo checked on that day. The model landscape ages fast: what you see here is a snapshot, not a leaderboard.

Size class

AllLaptopWorkstationServer

Built for

AllText & agentsCodeVisionEmbedding

Architecture

AllDenseMoE

12 models

• AlibabaQwen3.6-27B27BDenseText & agentsThe all-rounder for an agent on your own hardware: reliable tool calls, usable multilingual output, 256K of context.Details
• AlibabaQwen3.6-35B-A3B35B-A3BMoEText & agentsThe workhorse for long agent runs: it answers as fast as a 3B model because only 8 of 256 experts fire per token - yet all of it has to sit in memory.Details
• Mistral AIMinistral-3-3B-Instruct-25123BDenseText & agentsThe entry point without a GPU: roughly 2.5 GB at 4 bits, runs on an ordinary laptop.Details
• Googlegemma-4-12B-it12BDenseVisionA model that takes text, images and audio without a separate encoder - handy wherever your agent has to deal with screenshots, photos or voice notes.Details
• MetaLlama-4-Scout-17B-16E-Instruct109B-A17BMoEText & agentsThe model for very long inputs: Meta advertises up to 10 million tokens of context, 16 experts, 17B of them active.Details
• Mistral AIMistral-Small-4-119B-2603119B-A6.5BMoEText & agentsA server-class MoE that still computes cheaply: 128 experts, 4 of them per token, 6.5B active.Details
• AlibabaQwen3-Coder-30B-A3B-Instruct30B-A3BMoECodeThe code specialist for local developer agents: post-trained on repositories, patches and tool calls rather than chit-chat.Details
• AlibabaQwen3-VL-8B-Instruct8BDenseVisionA pure vision-language model at laptop size: read screenshots, capture forms, describe interfaces, follow video scenes.Details
• BAAIbge-m30.6BDenseEmbeddingNot a conversational partner but a translator into numbers: it turns chunks of text into 1024-dimensional vectors so that similar things can be found.Details
• Nous ResearchHermes-4.3-36B36BDenseText & agentsA fine-tune on someone else's foundation (ByteDance Seed-OSS-36B), built for tool calls, schema-conformant JSON and a switchable thinking mode.Details
• OpenAIgpt-oss-20b21B-A3.6BMoEText & agentsAn MoE shipped in 4 bits from the factory, which is why it starts at around 16 GB of memory.Details
• DeepSeekDeepSeek-V4-Pro1.6T-A49BMoEText & agentsThe top end of the open scale: 1.6 trillion parameters, 49B active, a million tokens of context.Details

Choosing a model is an engineering decision, not a religious war: task, budget and privacy decide who gets called - not this week's headline.

And your agent's architecture has priced that in from the start: model is a field in the request. The harness - conversation, role, tools, loop - stays exactly the same whether a 3B model on your laptop picks up or the biggest API model of the season.