LLM Models

xAI's first dedicated reasoning model (February 2025), trained on the Colossus supercluster and reported to outperform GPT-4o on academic benchmarks such as AIME 2025 and GPQA. Grok 3 serves a 131K-token context window via the xAI API and introduced Think and Big Brain reasoning modes.

xAI's frontier model at its July 2025 launch, positioned as its most capable reasoning model with native tool use and real-time search integration from X. Grok 4 features a 256K-token context window, and a Grok 4 Heavy multi-agent tier is offered to SuperGrok Heavy subscribers.

xAI's flagship Grok model (April 2026), recommended for chat and coding as its most intelligent and fastest model. Features a 1M-token context window, native video input (an xAI first), and tight integration with real-time search tools.

The efficient tier of DeepSeek's V4 generation (April 2026): a 284B total / 13B active MoE with a 1M-token context window and the same hybrid attention design as V4-Pro. MIT-licensed for low-cost, high-throughput deployments.

DeepSeek's frontier open-weights MoE model (April 2026): 1.6T total / 49B active parameters with a 1M-token context window and a hybrid attention mechanism (Compressed Sparse Attention + Heavily Compressed Attention) for efficient long context. MIT-licensed and competitive with top closed models on math, STEM, and coding.

Meta's long-context Llama 4 model (April 2025): a mixture-of-experts design with 17B active parameters across 16 experts (109B total) and an industry-leading 10M-token context window. Natively multimodal and open-weight under the Llama 4 Community License.

Meta's natively-multimodal Llama 4 mid-tier model (April 2025): a mixture-of-experts design with 17B active parameters across 128 experts (400B total) and a 1M-token context window. Open weights under the Llama 4 Community License with early-fusion multimodality.

LMSYS's open-source chatbot fine-tuned from LLaMA with supervised instruction fine-tuning on ~125K user-shared conversations from ShareGPT.com. Auto-regressive language model based on transformer architecture with 33B parameters. Evaluated with standard benchmarks, human preference studies, and LLM-as-a-judge (MT-Bench). Achieved 1091 ELO rating on Chatbot Arena. Scores: 62.1 Arc, 83 HellaSwag, 59.2 MMLU, 56.2 TruthfulQA, 77 WinoGrande, 13.7 GSM8K. Released as part of FastChat platform for training and evaluating large language models. Represents important early work in open-source chat model fine-tuning. Non-commercial license. Part of research leading to Chatbot Arena benchmark.

DISCONTINUED. MosaicML's (now Databricks) efficient 30B parameter decoder-style transformer pretrained from scratch on 1T tokens of English text and code. Features 8K token context window with support for extrapolation via ALiBi. Trained initially on 1T tokens using 2K sequences then additional 50B tokens of 8K sequences. First LLM trained on NVIDIA H100 GPUs. Outperforms originally published GPT-3 using ~1/6th the parameters. Competitive with LLaMa-30B and Falcon-40B on open-source benchmarks. Surpasses purpose-built models like StarCoder on HumanEval. Designed for easy deployment on single GPU - 1xA100-80GB (16-bit) or 1xA100-40GB (8-bit). Features FlashAttention for efficient inference. Released under Apache 2.0. Available in base, instruct, and chat variants. MosaicML no longer distributes the weights: the Hugging Face repository returns an authentication error and the mosaicml organisation no longer lists the MPT models.

Stability AI's enterprise audio model (September 2025) for high-quality sound and music production at scale. A generation breakthrough cuts inference from 50 steps to 8, producing three-minute tracks in seconds, and it supports text-to-audio, audio-to-audio, and audio inpainting on user files.

Cartesia's real-time text-to-speech model (late 2025), built on a state-space architecture for extremely low latency (~90ms model latency) with strong emotional expressiveness, multilingual support, and instant voice cloning from short samples - ideal for live voice agents.

Deepgram's speech-to-text model (January 2025), the replacement for Nova-2, delivering large word-error-rate reductions (54.3% streaming, 47.4% batch) versus competitors. Nova-3 is the first voice AI model with real-time multilingual transcription and self-serve keyterm customization without retraining.