latentbrief

Models

The frontier AI models, explained

The right pick is rarely the one topping a leaderboard — it's the one that fits your workload, budget and constraints. These are the models and the labs worth knowing.

San Francisco, CA

The safety-first AI lab that made alignment research a precondition for building. Claude models are known for disciplined instruction following, precise tool use, and a 200K context window that handles entire codebases in one pass.

Models, news & more →

San Francisco, CA

The lab that launched the current LLM era. GPT and o-series models anchor the widest developer ecosystem in the field — most tutorials, integrations, and third-party tooling start here.

Models, news & more →

Mountain View, CA

DeepMind and Google Brain, unified. The Gemini family brings native video and audio understanding and context windows up to 2M tokens — multimodal infrastructure at a scale no other lab matches.

Models, news & more →

Menlo Park, CA

Open weights by default. Meta releases Llama model weights publicly, giving teams full control to self-host, fine-tune and deploy frontier-grade models without API lock-in or per-token pricing.

Models, news & more →

Paris, France

The European challenger. Mistral builds efficient models with data residency and sovereignty as first-class concerns — clean function calling and GDPR-native infrastructure for teams where compliance is a hard requirement.

Models, news & more →

Hangzhou, China

The cost curve disruptor. DeepSeek challenged the assumption that frontier reasoning requires frontier pricing, then released the weights publicly — turning their advantage into a floor anyone can build on.

Models, news & more →

Hangzhou, China

China's open-weights powerhouse. The Qwen family spans 0.5B to 72B across text, vision, coding and math — with standout multilingual capability, especially in Chinese, that closed Western APIs can't match.

Models, news & more →

San Francisco, CA

Elon Musk's AI venture. Grok models differentiate on real-time X integration and inference-time compute — a unique angle for teams that need live web context beyond a training cutoff date.

Models, news & more →

Toronto, Canada

The enterprise retrieval specialist. Cohere focuses on retrieval-augmented generation and tool-calling rather than topping leaderboards. Command R+ is built for citation-accurate pipelines, with open weights so you're never locked in.

Models, news & more →

Legacy

Models we still cover for context but no longer recommend for new work.