Models

Choosing a model starts with understanding its capabilities, availability, and trade-offs.

Model categories

• General-purpose and code-specialized large language models
• Multimodal models
• Image-generation models, including text-to-image and image-to-image systems
• Decision or “System One” models, including calibrated-decision approaches

Model families and access

The landscape includes models from Anthropic (Claude), OpenAI (GPT), Google DeepMind (Gemini and Gemma), DeepSeek, Alibaba (Qwen), Meta (Llama and Muse), xAI (Grok), Moonshot (Kimi), Mistral, and Zhipu (GLM). Open-weight and closed models differ in how they can be inspected, hosted, adapted, and accessed.

Evaluation and training data

Benchmarks help compare models, but scores need context. Some benchmarks, such as MMLU and GSM8K, are nearing saturation; others, including SWE-bench variants, LiveBench, LMArena, GAIA, and BenchLM, provide different signals. Training data sources such as Common Crawl and FineWeb also shape model capabilities.

Abliteration

Refusal-removal techniques and fine-tunes, including Heretic and Ajax, can change how a model responds. Consider their impact on behavior and agent autonomy when selecting a model.