LLM Models

Ai2's fully open 32B model (March 2025) - the first fully open model (all data, code, weights, and training details public) to outperform GPT-3.5 Turbo and GPT-4o mini on a suite of multi-skill academic benchmarks. Trained on up to 6T tokens and post-trained with Tulu 3.1.

Meta's compact 3B open-weight model (September 2024), part of the Llama 3.2 edge collection, optimized for on-device summarization, rewriting, instruction following, and agentic retrieval. It offers a 128K context window and multilingual support at a size that runs on phones and laptops.

IBM's open 8B enterprise language model (2025), trained for general language, math, coding, and instruction following with a 128K context window. Released under Apache 2.0, the Granite family targets governed, on-prem enterprise deployments with strong tool use and RAG.

DISCONTINUED. AI21's open hybrid Mamba-Transformer (SSM) MoE model (August 2024), with 398B total parameters and 94B active. Its hybrid architecture delivers up to 2.5x faster long-context inference than comparable models, with a 256K context window suited to agentic and document-heavy workloads. AI21 deprecated it on May 6, 2025; its model card names AI21-Jamba-Large-1.7 as the successor and the current docs no longer list this version.

Tencent's open-source Transformer MoE (November 2024), among the largest open MoE models at 389B total parameters with 52B active, and a 256K context window. It matches or exceeds Llama 3.1 on many benchmarks and is released for research and commercial use.

Baidu's open-weight flagship text MoE (July 2025), part of the ERNIE 4.5 family open-sourced under Apache 2.0. The 300B-A47B model activates 47B parameters and delivers strong reasoning and Chinese/English performance, alongside multimodal and smaller variants from 0.3B to 424B.

Nous Research's flagship open model (August 2024), the first full-parameter fine-tune of Llama 3.1 405B. Hermes 3 is a steerable generalist with strong agentic capabilities, roleplay, reasoning, and long-context coherence, competitive with or exceeding Llama 3.1 Instruct.

DISCONTINUED. AI21's small hybrid Mamba-Transformer (SSM) model (August 2024), with 52B total parameters and 12B active, delivering long-context efficiency at small scale. It offers a 256K context window (up to ~140K on a single GPU) at very low cost, ideal for document-heavy and agentic workloads. AI21 deprecated it on May 6, 2025; its model card directs users to a newer version and the current docs no longer list it.

DISCONTINUED. Mistral's first multimodal model (September 2024), a 12B open-weight model with a dedicated vision encoder that understands images alongside text at native resolution. Released under Apache 2.0, it brought strong open multimodal performance to consumer-scale hardware. Mistral's own launch page now carries a deprecation banner stating the model is no longer maintained; the hosted version was deprecated on December 2, 2025 with retirement on December 31, 2025.

DISCONTINUED. Mistral and NVIDIA's jointly released 12B open model (July 2024), a drop-in upgrade to Mistral 7B with a 128K context window. Trained on multilingual and code data with a new Tekken tokenizer, it ships under Apache 2.0 in base and instruct versions plus an FP8 variant. Mistral deprecated the hosted model on May 22, 2026 with retirement on July 31, 2026, a date now passed; the Apache-2.0 weights remain on Hugging Face.

Cohere Labs' open multilingual model (December 2024) mastering 101 languages through data arbitrage, preference tuning, and model merging. Aya Expanse 32B outperforms Gemma 2, Qwen 2.5, and Llama 3.1 in its class on multilingual Arena-Hard-Auto, advancing the multilingual open frontier.

Cohere's open-weight 35B model for enterprise retrieval-augmented generation and tool use, with a 128K context window and strong multilingual support. Command R pairs with Cohere Rerank for high-precision RAG at a lower cost than the larger Command R+.