LLM Models

Cohere's enterprise flagship (March 2025), a 111B open-weight model with a 256K context window tuned for agentic tasks, RAG, tool use, and multilingual business workflows. It runs on just two GPUs and is ~150% more efficient at inference than Command R+.

Cohere's smallest and fastest enterprise model (December 2024), a 7B open-weight model optimized for retrieval-augmented generation, tool use, and agentic workflows. It runs on commodity and edge hardware with a 128K context window and multilingual support across 23 languages.

Cohere's enterprise-focused flagship model with 104B parameters and 128K context window. Released April 2024, updated August 2024. Optimized for advanced RAG with industry-leading grounded generation capabilities providing in-line citations to mitigate hallucinations. Trained with specific grounding capabilities - generates responses based on supplied document snippets including citation spans indicating information source. Supports 10 key business languages (trained on 23, evaluated on 10). Powerful Tool Use API for automating complex workflows. Beats GPT-4 Turbo on enterprise benchmarks. Available on Azure, OCI, and Cohere API. Pricing: $3 per 1M input, $15 per 1M output.

Meta's specialized code generation model with 34B parameters built on Llama 2 architecture. Released August 2023, trained on 500 billion tokens of code and code-related data from January-July 2023. Features extended 100,000 token context window for handling large codebases. Supports Python, C++, Java, PHP, TypeScript/JavaScript, C#, and Bash. Three variants: Base (code completion), Python (Python-specialized), Instruct (instruction-following). Achieves 53.7% HumanEval and 56.2% MBPP. Trained on Meta's Research Super Cluster requiring 400K GPU hours on A100-80GB hardware across all 9 CodeLlama models. Open weights enable customization and self-hosting.

Meta's largest open-weights Llama 2 model with 70B parameters. Released July 2023, trained on 2 trillion tokens with 40% more data than Llama 1. Features 4K context length (expanded from 2K in Llama 1) and Grouped-Query Attention for improved inference scalability. Supports both research and commercial use with permissive license. Pretrained on data through September 2022 with tuning data through July 2023. Global batch size of 4M tokens during training. Fine-tuned chat variant optimized for dialogue using supervised fine-tuning and RLHF. Widely adopted as foundation for custom models and applications. Superseded by Llama 3.1 70B (July 2024).

DISCONTINUED. Mistral AI's compact but powerful 7.3B parameter model released September 2023. First model from Mistral AI, setting new standards for efficiency and performance in its size class. Outperforms Llama 2 13B on all benchmarks and Llama 1 34B on many benchmarks despite being significantly smaller. Approaches CodeLlama 7B code performance while maintaining strong English capabilities. Features Grouped-Query Attention for faster inference and Sliding Window Attention (4096 tokens) for efficient long sequences. Apache 2.0 license. Deployable anywhere including locally, cloud (AWS/GCP/Azure), available on HuggingFace. Mistral has retired the hosted model - its docs mark it Status: Retired with a retirement date of March 30, 2025 - but the Apache-2.0 weights remain downloadable on Hugging Face.

DISCONTINUED. Mistral AI's groundbreaking Sparse Mixture of Experts model with 46.7B total parameters. Released December 2023, uses 13B active parameters per token while achieving performance of 6x larger models. Outperforms Llama 2 70B on most benchmarks with 6x faster inference. Matches or exceeds GPT-3.5 across all evaluated benchmarks. Features 32K context window with sliding window attention enabling theoretical 128K token span. Best open-weights chatbot model as of December 2023 per MT-Bench. Supports 5 languages (English, French, German, Spanish, Italian). Apache 2.0 license enables unrestricted commercial use. Mistral has retired the hosted model - its docs mark it Status: Retired with a retirement date of March 30, 2025 - but the Apache-2.0 weights remain downloadable on Hugging Face.

DISCONTINUED. Anthropic's balanced model striking ideal equilibrium between intelligence and speed for enterprise workloads. Released March 2024 alongside Opus and Haiku in Claude 3 family. Delivers 2x faster performance than Claude 2/2.1 with higher intelligence levels. Excels at tasks demanding rapid responses like knowledge retrieval or sales automation. Strong performance at lower cost compared to peers, engineered for high endurance in large-scale AI deployments. Features 200K context window and sophisticated vision capabilities processing photos, charts, graphs, technical diagrams. Positioned between Haiku (speed) and Opus (capability). Anthropic retired it on July 21, 2025; requests to retired models fail.

DEPRECATED. OpenAI's widely-adopted cost-effective model released March 2023. Optimized for chat via Chat Completions API with natural language and code generation capabilities. Features 16K context window (20 pages of text). Significantly cheaper than GPT-4 family while maintaining solid performance for most common use cases. Priced at $0.50 per 1M input tokens and $1.50 per 1M output (60% cheaper than previous pricing). Enhanced with better steerability, improved function calling, and extended context. Widely deployed in production applications globally. GPT-4o-mini now recommended upgrade for better performance at similar cost. OpenAI has deprecated it: access to the gpt-3.5-turbo snapshots is scheduled to shut down on October 23, 2026.

Stability AI's flagship open-source text-to-image generation model. Features 3.5B parameter base model with 6.6B parameter refiner in ensemble pipeline. Native 1024x1024 resolution (2x larger than SD 1.5) with improved generation for limbs, text, faces, and overall image quality. Uses dual CLIP networks (CLIP1 + CLIP2) for superior semantic understanding vs single CLIP. Achieves 89% prompt adherence vs SD 1.5's 71%. Supports image-to-image, inpainting, and outpainting workflows. Runs on consumer hardware (RTX 3060+ with 8GB VRAM).

Google's efficient mid-size open-weights model with 9B parameters. Released June-July 2024, trained on 8 trillion tokens primarily from web documents, code, and mathematics. Delivers class-leading performance, outperforming Llama 3 8B and other open models in its size category. Features Grouped-Query Attention, interleaved sliding window (4096 tokens) and global attention (8192 tokens). Trained with knowledge distillation from larger 27B model for improved quality. 8K context length. Optimized to run at incredible speed on range of hardware - gaming laptops, high-end desktops, cloud setups. Commercially-friendly license. Available via Google AI Studio, Kaggle, Hugging Face.

Google's largest open-weights model in Gemma 2 family with 27B parameters. Released June-July 2024, trained on 13 trillion tokens of web documents, code, and mathematics. Features interleaved local/global attention, logit soft-capping for stability, Grouped Query Attention for efficiency. Outperforms Qwen1.5 32B baseline and performs "only a few percent below" Llama 3 70B despite being significantly smaller. Optimized to run full precision on single Google Cloud TPU host, NVIDIA A100 80GB, or H100. 256K vocabulary with 4K context using RoPE. Commercially-friendly license. Available via Google AI Studio, Kaggle, Hugging Face.