Free AI Models

Cohere's advanced RAG-optimized model with 128K context and tool-use capabilities. Designed for enterprise retrieval-augmented generation workloads.

Databricks' open MoE model with fine-grained 16-expert architecture. Outperforms GPT-3.5 on coding and reasoning benchmarks with efficient 36B active parameters.

DeepSeek's first-generation reasoning model, rivaling OpenAI o1 on math, code, and reasoning benchmarks. Uses mixture-of-experts with 671B total parameters, 37B activated per token.

An efficient MoE language model with Multi-head Latent Attention (MLA) and DeepSeekMoE architecture. Competitive with GPT-4 at a fraction of the compute cost.

TII's Falcon 2 model with multilingual capabilities and optional vision adapter. Trained on 5T tokens with emphasis on Arabic and multilingual performance.

Google's open model built from the same research as Gemini. Uses knowledge distillation and a novel 2:1 interleaved attention architecture for efficient inference.

xAI's open-weight 314B MoE model. A massive mixture-of-8-experts architecture with 25% of parameters active per token. Rarely used for inference due to its size.

Meta's most capable open model in the Llama 3 family. Trained on 15T tokens with optimized data mixture. Strong performance across reasoning, coding, and general knowledge.

The efficient 8B variant of Meta's Llama 3 family. Punching well above its weight class, outperforming many larger models on key benchmarks.

Mistral's compact yet powerful 7B model with grouped-query attention and sliding window attention. Outperformed Llama 2 13B on most benchmarks at release.

Mistral's largest open MoE model. With 141B total parameters and a 64K context window, it competes with leading proprietary models on multitask benchmarks.

A high-quality sparse Mixture-of-Experts model matching or exceeding Llama 2 70B with significantly lower inference cost. Uses 8 experts with 2 active per token.

Microsoft's data-efficient model achieving strong results through carefully curated high-quality training data. Features a 128K context window despite its modest 14B size.

Alibaba's flagship open model. Features a massive 128K context window and strong performance across coding, math, and instruction-following. Outperforms Llama 3 70B on most benchmarks.

The efficient 7B variant of Qwen 2.5. Achieves remarkable per-parameter efficiency, often competing with models 3-4x its size on reasoning benchmarks.

01.AI's Yi 1.5 pretrained on 3.1T tokens of English and Chinese data with a 4K context window. Strong multilingual performance in a compact 34B package.