Allen AI API

Molmo 2 8B is an open vision-language model from the Allen Institute for AI (AI2), built on a Qwen3-8B language backbone with a SigLIP 2 vision encoder. It supports single images, multi-image inputs, and video clips. On its 11-benchmark image average, Molmo 2 8B leads all open-weight models in its class. It achieves 32.9% on video pointing versus 17% for Gemini 2.5 Pro, and tops open-weight scores across seven video benchmarks including Video-MME and MVBench. It also outperforms the original Molmo 72B on grounding tasks despite being far smaller. A strong choice for multimodal applications requiring precise spatial reasoning, visual grounding, or video understanding via API.

OLMo 3.1 32B Instruct is a fully open instruction-tuned language model from the Allen Institute for AI (AI2), designed for chat, tool use, and multi-turn dialogue at the 32B parameter scale. AI2 positions it as the most capable fully open 32B-scale instruct model, with strong performance on math (GSM8K, MATH), coding (HumanEval, MBPP+), and instruction-following (IFEval). It uses a hybrid attention architecture and maintains strong long-context retrieval performance (96.1 on RULER at 4K). Released under Apache 2.0 with full data and training transparency, it's a well-rounded choice for instruction-following, tool-augmented, or multi-turn chat applications.

OLMo 3.1 32B Think is the updated flagship reasoning model from the Allen Institute for AI (AI2), an improved successor to OLMo 3 32B Think with an additional 21 days of extended reinforcement learning training. The extended training yielded gains of 5+ points on AIME, 4+ points on ZebraLogic and IFEval, and 20+ points on IFBench over its predecessor. It supports a 64K context window and is licensed under Apache 2.0 with full training transparency. For API developers needing a high-performance open reasoning model for math, code, and complex instruction-following, OLMo 3.1 32B Think is AI2's most capable reasoning offering, competitive with Qwen 3 32B at the same scale.

OLMo 3 32B Think is a fully open reasoning model from the Allen Institute for AI (AI2), and the first fully open 32B thinking model to expose intermediate chain-of-thought reasoning traces. Trained on multi-step math, code, and general problem-solving tasks using a thinking SFT, DPO, and RLVR training flow, it is the strongest fully open reasoning model at the 32B scale — narrowing the gap to open-weight models like Qwen 3-32B-Think while trained on 6x fewer tokens. All training data, code, weights, and checkpoints are publicly available under Apache 2.0. Best suited for complex multi-step reasoning and mathematical problem solving via API.

OLMo 3 7B Instruct is a lightweight, fully open instruction-tuned chat model from the Allen Institute for AI (AI2), designed for instruction-following, question-answering, and multi-turn conversational dialogue. Among 7B-scale models, it is competitive with Qwen 2.5 and Gemma 3 equivalents, and represents a clear step up from Llama 3.1 8B in instruction-following quality. It supports a 66K token context window with a knowledge cutoff of December 2024. Released under a fully open license with complete training weights and data publicly available, it's well-suited for cost-efficient API usage where a capable small model is preferred.

OLMo 3 7B Think is an efficient reasoning model from the Allen Institute for AI (AI2), purpose-built for multi-step problem solving in math, coding, and general analytical tasks. Trained using a thinking SFT, thinking DPO, and RLVR pipeline, it generates structured chain-of-thought reasoning traces. On math benchmarks it matches Qwen 3 8B on MATH and comes within a few points on AIME 2024 and 2025. On coding, it leads similarly-sized models on HumanEvalPlus. Fully open under a permissive license, it is the most capable fully open reasoning option at the 7B scale — ideal for API use cases requiring strong reasoning at a small model footprint.

OLMo 2 32B Instruct is a fully open, 32-billion-parameter instruction-tuned language model from the Allen Institute for AI (AI2), post-trained using supervised fine-tuning, DPO, and RLVR. It is the first fully open model to outperform both GPT-3.5 Turbo and GPT-4o mini across popular multi-skill academic benchmarks including GSM8K, MATH, and IFEval. The model supports a 128K token context window and targets math reasoning, instruction-following, and general chat. Released under Apache 2.0 with full transparency into training data, code, and weights, it's a strong choice for developers who need a capable, commercially permissive instruction model.