DeepSeek V3
685B MoE model trained for $5.5M — remarkable efficiency
Verdict
Proved you can train frontier models affordably. 685B MoE with FP8 training on 2048 H800s. Strong general capabilities. Best for teams wanting frontier quality with efficient self-hosting.
Other Text Generation & Reasoning
- Claude 3.5 SonnetProduction
Anthropic's proven workhorse with 200K context and computer use
- Claude 4 SonnetProduction
Anthropic's latest — best-in-class coding, reasoning, and safety
- Command R+Stable
Cohere's enterprise model optimised for RAG and tool use
- DeepSeek R1Production
Open-weight chain-of-thought reasoning rivalling GPT-o1
- GLM-5Experimental
Zhipu AI's 754B frontier model from China's leading AI lab
- GPT-3.5 TurboSuperseded
OpenAI's original cost-efficient chat model — now superseded