Solutions for
Please select
Try the Web App
ki-modelle

LLM Model Families at a Glance

Redaktion ·

Why this overview

Anyone setting up an AI project in 2026 is spoiled for choice: Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, Alibaba, xAI, Moonshot, Zhipu — every vendor maintains at least one, often several model families. On top of that, there are open weights, closed APIs, regional specialists, reasoning variants and multimodality. „Which model do I take?” is not a one-sentence question — but it becomes tractable once you read the landscape as a playing field with clear poles: Western vs. Chinese, closed vs. open, frontier vs. mid-tier vs. low-tier, generalist vs. specialist.

This article sorts the most important families along exactly these axes, describes typical strengths and weaknesses, and gives you a pragmatic rule of thumb at the end for when which family makes sense. Detail topics like pricing, local models or fine-tuning are standalone articles — here it’s about orientation.

The Western frontier vendors

Anthropic — Claude

Anthropic is the company behind the Claude family. The current lineup has three tiers: Opus as the top model for complex reasoning and coding, Sonnet as the workhorse for most production applications, Haiku as the cheap, fast variant for classification, routing and short answers. Across the industry Claude is regarded as the model with the strongest results in agentic coding, tool use (function calling) and understanding long, mixed contexts (1 M tokens on Sonnet). Characteristic traits: a rather restrained, precise tone of voice and comparatively strict behaviour on safety prompts.

Typical fields: coding agents, RAG systems with long documents, customer-service bots with tool use, pipeline stages where accuracy matters. Weakness: classic multimodality (image generation, audio output) is not the focus — for that there are specialised vendors or Google.

OpenAI — GPT and the o series

OpenAI runs two families in parallel: the GPT line (GPT-5.5, GPT-5.5-mini, GPT-Nano) as generalist language models and the o series (o4, o4-mini) as reasoning specialists with a longer internal thinking phase before answering. GPT models have the broadest market reach, the largest ecosystem of third-party tools, the most integrations into office software, and the most mature multimodality (image, audio, video).

Typical fields: general-purpose assistants, multimodal applications (speech-to-text, image analysis), standard chatbots with high user expectations, applications needing broad tool integration. Weakness: in pure coding and very long, complex tool chains Claude is often ahead. Reasoning models from the o series can become disproportionately expensive on simple questions.

Google — Gemini and Gemma

Google offers two related lines: Gemini (Pro, Flash, Flash-Lite) as closed API models and Gemma as open weights for self-hosting. Gemini’s unique selling point is native multimodality from the start — images, audio and video are not bolted on later but part of training. Add to that an extremely long context (2 M tokens on Pro) and tight integration with Google services (search grounding, Workspace).

Typical fields: multimodal pipelines (video understanding, PDF OCR), applications that need live web data via search grounding, processing very large data chunks at once. Gemma is the entry point for everyone who prefers a Google flavour but has to host locally. Weakness: tool use is less mature than Claude or GPT, and consistency across multiple turns is considered weaker.

Meta — Llama

Meta publishes the Llama family as open weights. The current Llama-4 generation (Scout, Maverick, Behemoth) combines a mixture-of-experts architecture with long context windows and lands close to Western frontier models on many benchmarks. Llama is the default model for any self-hosting setup that needs a Western licence, and the base for countless fine-tunes (Code-Llama, Llama-Guard, vertical industry forks).

Typical fields: on-premise deployments with data-protection requirements, fine-tuning for vertical markets, edge devices with Llama-Nano variants. Weakness: out of the box Llama is often less „helpful” than Claude or GPT — models benefit a lot from fine-tuning and downstream instruction shaping. Anyone going to production without these steps should prefer Sonnet/GPT.

The European specialist

Mistral

Mistral AI is the most prominent European vendor. The portfolio mixes open models (Mistral Small, Mistral 7B, Codestral) with proprietary API models (Mistral Large, Mistral Medium). Strengths: low token costs, good performance on European languages beyond English, a clear compliance story (Paris-based, GDPR-compliant with EU data residency).

Typical fields: multilingual applications focused on French, German, Italian, Spanish; government and healthcare projects with strict data-protection requirements; self-hosting setups with small models (7B–22B). Weakness: at the very top tier Mistral Large does not quite reach the US and Chinese frontier — for reasoning-heavy tasks other families are stronger.

The Chinese vendors

In two years the Chinese model landscape has gone from outsider to frontier status. Four families are particularly relevant.

DeepSeek

DeepSeek became world-famous in 2025 with the V3 release: a mixture-of-experts model with GPT-4-level performance at a fraction of the training and inference cost. The current V4 line and the R1 reasoning model continue the trend. DeepSeek weights are openly available and hosted by many inference providers (Together, Fireworks, DeepInfra) — including in Western data centres.

Typical fields: cost-sensitive high-volume workloads, reasoning tasks under budget pressure, research and experimentation projects. Note: anyone using the DeepSeek API directly in China should be aware of the data implications — Western hosting providers sidestep that.

Qwen (Alibaba)

Alibaba’s Qwen family is the broadest Chinese model line: dense models, MoE models, vision-language variants, coding-specific versions. Qwen regularly hits top-3 results on benchmarks and is by far the strongest open option for Mandarin and other Asian languages.

Typical fields: applications with an Asian target audience, cost-sensitive multilingual pipelines, self-hosting with high variant diversity. Weakness: weaker tool use and agent integration than Western frontier models.

Moonshot AI — Kimi

Moonshot AI is the startup behind Kimi, a model with the longest practically usable context window today (several million tokens). Kimi is heavily used in Chinese consumer apps and is considered particularly strong at processing very long documents.

Typical fields: long-document analysis, research assistants, applications with an Asian market focus.

Zhipu AI — GLM

Zhipu AI (brand name Z.ai) builds the GLM line. Strengths are in academic applications, long reasoning chains and domain specialisations for the Chinese market.

The challengers

xAI — Grok

xAI (Elon Musk’s company) develops the Grok family. Grok-3 and Grok-4 are competitive on reasoning, benefit from live data from the X/Twitter stream and position themselves deliberately as „less restrained” than Claude or GPT. Adoption outside the X platform is limited, and the API ecosystem is still young.

Selection heuristic

If you have to make a first pick without research, this rule of thumb usually works:

  • Coding agent, tool use, long pipelines: Claude Sonnet as default, Opus for the trickiest steps.
  • Multimodal application, image and video understanding: Gemini Pro.
  • General-purpose chatbot with broad ecosystem needs: GPT-5.5 or GPT-5.5-mini.
  • Volume workload on a tight budget: DeepSeek V4 or Gemini Flash, depending on the task.
  • Self-hosting with a Western licence: Llama 4 or Mistral.
  • GDPR-strict application, EU hosting: Mistral Large, alternatively Claude/GPT/Gemini via EU regions.
  • Asian market, Mandarin as primary language: Qwen or Kimi.
  • Local dev model on laptop or workstation: Llama 3.x in 8B quantisation, Gemma 3, or a small Mistral.

What stays true, what changes fast

Model names and tier prices move on a quarterly rhythm. The axes — who builds which type, which family targets which strength — are surprisingly stable. Anthropic stays the coding and safety focus, OpenAI the multimodal and ecosystem focus, Google the long-context and multimodal focus, Meta the open-weights anchor, Mistral the EU anchor, DeepSeek/Qwen the cost- and open-source-driven Asians.

If you’re up against a concrete model choice, two more articles will help: AI pricing made simple for the economic side and Measuring LLM quality for the question of how to test models against your own use case. If you can’t or don’t want to rely on cloud APIs, the hosting options are covered in Running local LLMs.

Themenuebersicht

KI-ModelleClaudeGPTGeminiLlamaMistralDeepSeekQwenGrokGemma