Olmo 3

Abstract: We introduce Olmo 3, a family of state-of-the-art, fully-open LLMs at the 7B and 32B parameter scales. Olmo 3 model construction targets long-context reasoning, function calling, coding, instruction following, general chat, and knowledge recall. This release includes the entire model flow, i.e., the full lifecycle of the family of models, including every stage, checkpoint, data point, and dependency used to build it. Our flagship model, Olmo 3 Think 32B, is the strongest fully-open thinking model released to-date.

• The paper presents Olmo 3's complete open model flow, releasing training data, code, and intermediate checkpoints to enhance transparency and reproducibility.
• It employs advanced data curation, a scalable transformer with sliding window attention, and innovative RL techniques to boost stepwise reasoning and efficiency.
• Key results demonstrate state-of-the-art performance in math, code, and logic benchmarks while reducing training tokens by 6× compared to comparable models.

Olmo 3: A Fully-Open Model Flow for Scalable Reasoning and RL-Finetuning

Olmo 3 introduces a fully-open family of LLMs at 7B and 32B parameter scales, providing both the final model checkpoints and, uniquely, the entire "model flow"—including all training data, code, and intermediate checkpoints across pretraining, midtraining, long-context extension, and post-training stages. This comprehensive release paradigm supports deep reproducibility, transparency, and research into model development, fine-grained data influence, and RL adaptation. Olmo 3’s architectural and data innovations yield state-of-the-art performance for open models in stepwise reasoning, code, math, function calling, and general chat, outperforming leading alternatives such as Stanford Marin, LLM360, and Apertus in key metrics, and narrowing the gap with best-in-class closed open-weight models.

---

Model Flow Paradigm: Full Pipeline Openness

Olmo 3’s central contribution is the release of the full model workflow: every version and artifact from base pretraining datasets to post-trained models. This enables:

• Deep inspection and intervention: Researchers can analyze the effect of each stage, ablate data or checkpoints, and trace reasoning chains to original data.
• Replicability and transparency: All code (training, data processing, evaluation) and hashed data are available, establishing a new openness standard.
• RL and data contamination studies: RL from base (RL-Zero) is supported by open, decontaminated RL datasets, addressing previously confounded benchmarks.

  Figure 1: The model flow encompasses training data, code, and intermediate checkpoints for all stages of development.

---

Figure 2: Development is divided into major stages of base model training and post-training, each with their own curation and optimization recipes.

---

Data Pipeline Innovations

Olmo 3’s pretraining and midtraining data curation leverages several algorithmic advances:

• Global deduplication at trillion-token scale: Builds a massive, minimally-redundant, high-quality web and academic (olmOCR-processed PDF) corpus.
• Quality-aware upsampling and topic balancing: Uses topic/quality classifiers with swarm-based mixing to optimize for downstream capabilities (e.g., upweighting STEM domains for math and code gains).

  Figure 3: Data curation flow for pretraining data sources in Dolma 3 Mix.

---

Figure 4: Quality-aware upsampling curve compared to flat upsampling—showing more repetition for top-quality data, while discarding lower quality buckets.

---

Architectural Details and Scaling

• Decoder-only architecture: Transformer with 8192 token context window (pretrain), extended to 65,536 tokens.
• Sliding Window Attention (SWA): Used for scalable long-context pretraining, with final full-attention in upper layers.
• Long-context extension: Incorporates best-fit document packing, intra-document masking, and synthetic aggregation data (inspired by CLIPPER), yielding high RULER/HELMET scores at both 7B and 32B scales.

---

Figure 5: Learning rate schedule and loss for Olmo 3 7B, with cosine schedule and staged annealing.

---

Rigorous Evaluation Design

The OlmoBaseEval benchmark suite covers:

• Task clusters via hierarchical clustering: Aggregates signal across thematically-related benchmarks to reduce noise and maximize sensitivity to data/model interventions.
• Scaling and signal-to-noise analysis: Tracks emergence and saturation regions across sizes; removes benchmarks with low decision utility.
• Decontamination: Systematic n-gram/semantic filtering ensures no contamination between data and evaluation sets.

  Figure 6: OlmoBaseEval signal-to-noise analysis, with aggregation and hyperparameter tuning for robust signal.

---

Specialized Post-Training: Reasoning, Efficiency, RL

Olmo 3 Think

• Step-by-step reasoning models: Trained with curated thinking-trace datasets, high-quality contrastive pairs (for DPO), and RL with verifiable rewards (multi-domain).
• Strongest fully-open thinking model at 32B: On math, code, and logic reasoning benchmarks, matches or exceeds Marin/Apertus/LLM360; approaches Qwen 3 and Gemma 3 despite using 6× fewer tokens.

Olmo 3 Instruct

• Direct, latency-optimized model: Instruction-following and function-calling without intermediate reasoning traces, for practical chat and API use cases.

RL-Zero: Fully Open RL from Base

• Decontaminated RLVR setup: Provides open data, code, and benchmarks for clean RL training from base—critical for controlling for contamination/benchmark inflation in RL-based capability studies.

Figure 7: RL-Zero runs in diverse domains. RL from base model shows measurable improvements on math/coding/IF benchmarks, with isolated, decontaminated data tracing.

---

Reinforcement Learning: Algorithmic and Systems Advances

A suite of RL algorithm refinements and custom infrastructure achieves both sample efficiency and stability:

• Improved GRPO-based objective: Token-level loss, truncated importance sampling, "clip-higher", no KL term, group-advantaged filtering—all boosting convergence and sample utility.
• Verifiable and LLM-judge rewards: For math/code/instruction domains, RL uses automatic verifiers and tightly benchmarked LLM-judges.
• Continuous batching and inflight parameter updates: Overlap generation and training for high GPU utilization, reducing costly bottlenecks from autoregressive rollout generation.

Figure 8: Distributed reinforcement learning infrastructure, decoupling rollout generation and training for large-scale RL with long contexts.

---

Key Empirical Results

Post-Training Benchmarks: Olmo 3 Think 32B (Selected)

Benchmark	Olmo 3 Think 32B	Qwen 3 32B	Gemma 3 27B
MATH	96.2	95.4	87.4
AIME 2024	80.6	80.8	28.9
HumanEvalPlus	91.5	91.2	79.2
IFBench	68.1	37.3	31.3
ZebraLogic	80.1	88.3	24.8
AlpacaEval 2 LC	69.1	75.6	65.5
PopQA	30.9	30.7	30.2

Notably, Olmo 3 Think 32B reduces training FLOPs (6× fewer tokens) while matching or closely tailing best open-weight models. It sets a new standard among fully-open releases.

---

Practical and Theoretical Implications

• Reproducibility and interpretability: The full pipeline openness makes Olmo 3 ideal for studies of training stage, data, architecture, and RL influence on model capabilities and behaviors.
• RL as a scientific tool: With open RL-from-base data and code, contamination and spurious reward studies (e.g., non-improvement on spurious reward, as in [shao2025spuriousrewardsrethinkingtraining]) validate RL algorithmic progress and ground future research.
• Benchmarking RLHF/RLVR recipes: The open model flow allows for head-to-head comparisons under perfectly controlled setups (i.e., same base data, different RL algorithms).

---

Prospects for Future Development

Olmo 3’s paradigm of full model flow openness—combining large-scale deduplicated, quality-controlled data, strong capability-driven curation, full life-cycle public artifacts, and robust RL—lays the groundwork for:

• Data-centric RL studies: Isolating the effect of novel midtraining/curation recipes on RL-adapted capabilities.
• Post-training research: Ablations across SFT, DPO, RL for reasoning, function use, and chat—critical for understanding diminishing returns and oversaturation, and for optimizing with minimal resource use.
• Efficient large LLM development: Systemic advances (e.g., upsampling, batching, inference-speed RL infrastructure) can be adopted in both open and closed model pipelines for scalable, transparent progress.

---

Conclusion

Olmo 3 provides an unprecedented level of openness in large-scale model release, empowering the research community to study, adapt, and deeply understand the full pipeline of state-of-the-art LLMs. By setting new benchmarks for both transparency and capability—especially in reasoning and RL adaptation—it both closes key capability gaps and unlocks fine-grained, data- and method-driven research previously infeasible with closed or partially-open model releases. This work marks a significant step toward principled, empirical foundation model science and engineering.

---