Allen Institute for AI

AstaBrief

Allen Institute for AI

AstaBrief is Ai2’s 8B scientific report-generation model, trained from Qwen3-8B with citation-focused supervised fine-tuning and direct preference optimization. It turns a research question and retrieved literature excerpts into a complete cited report in a single pass, and powers Asta’s Fast report mode. Ai2 releases the weights, training data, and a workflow for using local PDFs. The reported latency advantage evaluates the complete report pipeline rather than the checkpoint alone. Citation grounding is improved but does not guarantee that a synthesis preserves every study’s scope or conclusions.

EMO

Allen Institute for AI

EMO is Ai2’s open mixture-of-experts language model trained so reusable domain modules emerge without predefined semantic labels. Its 128 routed experts activate eight per token, giving 14B total and about 1B active parameters. A document-level routing objective encourages coherent expert groups that can be selected for particular tasks. The main checkpoint is trained on one trillion tokens with additional annealing. Ai2 releases weights, matched baselines, and training code for modularity research. Results for reduced expert subsets depend on the selection method and evaluation task.

BAR

Allen Institute for AI

BAR, or Branch-Adapt-Route, is Ai2’s released modular language-model family built from an Olmo 2 7B anchor and independently trained capability experts. Math, coding, tool-use, and safety experts are combined through a router into a final five-expert mixture, enabling capability updates without retraining the full model. The public checkpoints include the original anchor, intermediate two-expert combinations, and final combined model as variants. This is a concrete weight release supporting a post-training method, rather than a framework-only announcement. Expert merging and routing must be evaluated for interactions and task quality.

MolmoWeb

Allen Institute for AI

MolmoWeb is a released vision-language model trained for browser tasks from screenshots and natural-language instructions. Its 4B and 8B weights are distinct from the browser harness: the checkpoints generate textual reasoning, browser actions, and task-completion responses. Training uses MolmoWebMix, an open collection of synthetic and human browsing trajectories. Native and Transformers-compatible checkpoints are serving variants; compatibility fixes after launch do not establish a new model generation. Completing a web task depends on the external browser, observation/action interface, and execution permissions.

MolmoPoint

Allen Institute for AI

MolmoPoint is Ai2’s vision-language release with a token-based pointing mechanism that selects image regions directly from visual features rather than spelling out coordinates as ordinary text. It combines natural-language image and video understanding with spatial grounding, video pointing, and tracking. The 8B model uses Qwen3-8B and SigLIP; related 4B video and 8B GUI checkpoints specialize the same architecture. Weights, training code, and data support research on reliable visual interaction. Available implementations differ in which grounding functions they expose, and visual selections can still be inaccurate.

Olmo Hybrid

Allen Institute for AI

Olmo Hybrid is a fully open 7B language model that combines Transformer attention with gated DeltaNet linear recurrent layers. Three quarters of its sequence-mixing layers use gated DeltaNet, reducing long-context state costs while retaining attention in the remaining layers. Ai2 releases base, instruction, and thinking checkpoints alongside data and training artifacts. The model studies scaling and expressivity compared with Olmo 3 under controlled training, and provides a practical platform for hybrid-model experiments. Throughput comparisons depend on sequence length and hardware rather than representing a fixed speedup for every task.

SERA

Allen Institute for AI

SERA is Ai2’s first Open Coding Agents model family, trained to make repository-level software changes. Soft Verified Generation constructs supervised agent trajectories without requiring every trace to pass strict execution verification, reducing training cost while preserving useful development behavior. Released 8B, 14B, and 32B checkpoints include general and repository-adapted variants. The main 32B checkpoint derives from Qwen3-32B and uses a 32K task context. SWE-bench outcomes depend on the coding harness and environment. The research models do not include general safety filtering and require oversight for executing generated code.

Olmo 3.1

Allen Institute for AI

Olmo 3.1 is a fully open Ai2 language-model family. This significant 32B update extends the Olmo 3 thinking model’s reinforcement-learning run and adds a large instruction tier for chat, tool use, and multi-turn dialogue. Think and Instruct are complementary public checkpoints, with their own post-training objectives. The December 12 dated update in the original Olmo 3 blog establishes release. Apache-2.0 weights permit adaptation and deployment under the published terms. Outputs can hallucinate, and smaller tiers or different training stages should not be assumed to share identical capabilities.

Molmo 2

Allen Institute for AI

Molmo 2 is an open-weight Ai2 multimodal language-model family. Molmo 2 extends the family to video reasoning, pointing, and tracking with open training data and code. The 8B checkpoint uses Qwen3-8B with SigLIP 2; a 7B Olmo-backed tier and smaller 4B variants share the release’s vision-language approach. Later ER and VideoPoint specializations remain related variants. It produces natural-language explanations from visual evidence and supports research on grounded interaction. Images or videos can be misinterpreted, particularly at fine resolution or across difficult temporal sequences.

Olmo 3

Allen Institute for AI

Olmo 3 is a fully open Ai2 language-model family. Olmo 3 exposes a complete model flow from pretraining through instruction following and reasoning. The 7B and 32B base tiers support Instruct and Think post-training branches, along with RL-Zero research variants. Training data, intermediate weights, and reproducibility tools make the model useful for controlled studies as well as local assistants. Apache-2.0 weights permit adaptation and deployment under the published terms. Outputs can hallucinate, and smaller tiers or different training stages should not be assumed to share identical capabilities.

DR Tulu

Allen Institute for AI

DR Tulu is Ai2’s open deep-research language model and end-to-end training recipe for producing long-form grounded reports. The released 8B checkpoint builds on Qwen3-8B with supervised training and Reinforcement Learning with Evolving Rubrics. It learns to search, gather evidence, and synthesize reports through an external research harness. Ai2 releases checkpoints under the RL ReSearch organization, along with training data, tool stack, and evaluation artifacts. Report-quality scores measure the complete model-plus-search system and are not general standalone factual accuracy.

FlexOlmo

Allen Institute for AI

FlexOlmo is Ai2’s modular open language model for combining independently trained data-domain experts without pooling their underlying data. Seven expert branches cover public-mix, news, mathematics, code, academic writing, creative writing, and Reddit. Its 33B mixture-of-experts model demonstrates selective inclusion and exclusion of datasets through expert composition and routing. Ai2 releases weights and methods to study flexible data collaboration and governance. The expert label describes a training domain rather than a guaranteed factual specialty, and removing an expert can change behavior on other domains.

OLMo 2

Allen Institute for AI

OLMo 2 is a fully open Ai2 language-model family. The second-generation fully open family improves the training mix, architecture, and post-training process. Initial 7B and 13B tiers were followed by 32B and 1B releases in 2025. Data, code, and training stages are released for reproducibility; base, instruct, and routine size additions are variants. Apache-2.0 weights permit adaptation and deployment under the published terms. Outputs can hallucinate, and smaller tiers or different training stages should not be assumed to share identical capabilities.

Tülu 3

Allen Institute for AI

Tülu 3 is Ai2’s open post-training model family built on Meta Llama 3.1. It combines curated supervised instruction data, direct preference optimization, and reinforcement learning with verifiable rewards to improve chat, mathematical reasoning, instruction following, and tool use. The 8B and 70B releases were followed by a 405B tier and a 3.1 post-training refresh. Training data, scripts, and intermediate stages are public, but underlying Llama weight-license conditions still apply. This family page groups the routine size and post-training-stage checkpoints.

Molmo

Allen Institute for AI

Molmo is an open-weight Ai2 multimodal language-model family. The initial family combines vision encoders and pretrained language backbones with human-annotated PixMo image data, avoiding reliance on proprietary vision-language distillation. It answers visual questions, counts objects, and grounds responses with pointing. Backbones and license details vary across sizes, while the released vision-language training pipeline is open. It produces natural-language explanations from visual evidence and supports research on grounded interaction. Images or videos can be misinterpreted, particularly at fine resolution or across difficult temporal sequences.

OLMoE

Allen Institute for AI

OLMoE is a fully open Ai2 language-model family. Ai2’s openly trained sparse language model studies efficient routing and expert specialization. It provides a seven-billion-parameter weight footprint with approximately one billion parameters used per token. The initial base and instruction models are accompanied by the full training mixture, code, logs, and analysis; January 2025 instruction refreshes remain variants. Apache-2.0 weights permit adaptation and deployment under the published terms. Outputs can hallucinate, and smaller tiers or different training stages should not be assumed to share identical capabilities.

OLMo

Allen Institute for AI

OLMo is a fully open Ai2 language-model family. The original Open Language Model project releases model weights, Dolma training data, code, logs, and intermediate checkpoints to make the full language-model development process inspectable. Base models are designed for continuation rather than polished chat; later instruction and April/July data refreshes belong to this family. Apache-2.0 weights permit adaptation and deployment under the published terms. Outputs can hallucinate, and smaller tiers or different training stages should not be assumed to share identical capabilities.