Olmo 3 — part 2/19

Cleaned-text code points [15710, 29482). Bibliography omitted here; source.md includes it.

• Olmo 3 RL-Zero (Section § 6 ) To date, all leading open RLVR benchmarks and algorithms train on top of open-weight models that do not reveal their pretraining or mid-training data ( Chu et al., 2025 ; Yang et al., 2025a ) . This limits the community’s ability to study the role of pretraining data on RLVR performance. It can lead to myriad issues with benchmark evaluations being contaminated, e.g., mid-training data containing the evaluation, which makes spurious rewards as effective as true reward ( Shao et al., 2025b ; Wu et al., 2025c ) or improvements from fixing prompt templates outweighing the improvements from RL ( Liu et al., 2025b ) . ∘ \circ We therefore release a fully open dataset Dolci RL-Zero , an algorithmic RL zero setup for Olmo 3 , and open-source OlmoRL code to enable clear benchmarking in the RL research community. We perform RLVR from Olmo 3 Base over four benchmarking domains to create the Olmo 3 RL-Zero family: math, code, precise instruction following (IF) and a general mix. In all cases, we further decontaminate Dolci RL-Zero from pretraining and midtraining data to guarantee our setup carefully studies the effect of RLVR without data leakage confounding our conclusions.

Fully-Open Models Open-weight Models OLMo 3.1 32B Think OLMo 2 Instruct 32B Apertus Instruct 70B LLM360 K2-V2 Instruct 70B Qwen 3 32B Qwen 3 VL 32B Think Qwen 2.5 32B Gemma 3 27B Gemma 2 27B DS-R1 32B Math MATH 96.2 49.2 36.2 94.5 95.4 96.7 80.2 87.4 51.5 92.6 AIME 2024 80.6 4.6 0.3 78.4 80.8 86.3 15.7 28.9 4.7 70.3 AIME 2025 78.1 0.9 0.1 70.3 70.9 78.8 13.4 22.9 0.9 56.3 OMEGA 53.4 9.8 5.6 46.1 47.7 50.8 19.2 24.0 9.1 38.9 Reasoning BigBenchHard 88.6 65.6 57.0 87.6 90.6 91.1 80.9 82.4 66.0 89.7 ZebraLogic 80.1 13.3 9.0 79.2 88.3 96.1 24.1 24.8 17.2 69.4 AGI Eval English 89.2 68.4 61.7 89.6 90.0 92.2 78.9 76.9 70.9 88.1 Coding HumanEvalPlus 91.5 44.4 42.9 88.0 91.2 90.6 82.6 79.2 67.5 92.3 MBPP+ 68.3 49.0 45.8 66.0 70.6 66.2 66.6 65.7 61.2 70.1 LiveCodeBench v3 83.3 10.6 9.7 78.4 90.2 84.8 49.9 39.0 28.7 79.5 IF IFEval 93.8 85.8 70.4 85.3 86.5 85.5 81.9 85.4 62.1 78.7 IFBench 68.1 36.4 26.0 57.7 37.3 55.1 36.7 31.3 27.8 23.8 Knowledge & QA MMLU 86.4 77.1 70.2 88.4 88.8 90.1 84.6 74.6 76.1 88.0 PopQA 30.9 37.2 33.6 32.2 30.7 32.2 28.0 30.2 30.4 26.7 GPQA 57.5 36.4 27.9 64.0 67.3 67.4 44.6 45.0 39.9 61.8 Chat AlpacaEval 2 LC 69.1 38.0 19.9 - 75.6 80.9 81.9 65.5 39.8 26.2

### 2.3 Results

Table 1 demonstrates a snapshot of our evaluation for Olmo 3 Think compared to other open-weight and fully-open models. To the best of our knowledge, Olmo 3 Think is the strongest fully-open thinking model to date. It is better than Qwen2.5-Instruct, Gemma 2 and 3 27B, DeepSeek R1, and Distilled Qwen 32B; it is also close to Qwen 3 and Qwen 3 VL 32B models, narrowing the gap to the best open-weight models of similar scale while training on roughly 6x fewer tokens.

For more details and results of other models along our Olmo 3 model flow, refer to the quick links below.

• Olmo 3 Base Section § 3.7 for detailed evaluation discussion. Table 2 (32B) and Table 3 (7B) for main results. Table 12 for long context evaluations. Table 13 for pretraining vs midtraining vs long-context extension stages.

• Olmo 3 Think Section § 4.1 for detailed evaluation discussion. Table 14 (32B) and Table 15 (7B) for main results, including SFT vs DPO vs RL stages.

• Olmo 3 Instruct Section § 5.1 for detailed evaluation discussion. Table 25 (32B) and Table 26 (7B) for main results, including SFT vs DPO vs RL stages.

### 2.4 Costs

The cost of training large models is often reported as a single dollar figure, typically by converting GPU-hours at market rates to dollars, such as $5.576M in H800-hours for DeepSeek V3 ( DeepSeek-AI et al., 2025 ) . To provide a more representative view of the resources required to train Olmo 3 32B, we instead report the wall-clock time that elapses during training.

In total, approximately 56 days elapsed from the start of training to the evaluation of the Olmo 3 Think 32B checkpoint, on a cluster with 1024 H100 GPUs dedicated to Olmo 3 . The 32B 3.1 Think and Instruct checkpoints were trained after this time period. This training time is largely a reflection of applying our best recipe to the model 2 2 2 The recipe was developed on 7B or smaller and applied to 32B rapidly. , and does not include any substantial modifications or research ideas that could expand the timeline substantially. At a price of $2/H100 hour, this would cost $2.75M. Runtime breakdown is as follows:

• Pretraining: ∼ \mathbf{\sim} 47 days (including midtraining and long-context stages) The initial pretraining phase on 5.5T tokens took about 9.5 days on 512 GPUs, followed by an additional 35 days on 1024 GPUs. These durations include all crash resumptions and other engineering concerns that kept us from running at full speed. Midtraining consisted of two parallel runs on 512 GPUs each, covering 100B tokens per run, followed by model merging and evaluations to decide on final checkpoints, taking about 1.5 days in total. Long-context extension was executed as a single run on 1024 GPUs; the full long-context stage—including training and all associated merges and evaluations—added approximately one additional day.

• Post-training: ∼ \mathbf{\sim} 9 days (SFT, DPO, and RL) Post-training follows a different operational pattern in which we run each stage multiple times, sweeping over learning rates and other hyperparameters. The theory for post-training, particularly, RL, is less developed, so we have to run multiple experiments to identify the optimal hyperparameters for a given base model. We hope to address this in future work. During post-training, checkpoint evaluation consumes a larger proportion of compute resources, in part due to long generations from reasoning models on core benchmarks. For SFT, we swept over four candidate learning rates, on 256 GPUs each, in parallel for 36 hours. Then approximately 12 hours was spent on evaluation, merging, and checkpoint confirmation, totaling approximately two days. DPO training takes less time per run (about 18 hours for a full learning-rate sweep on 64 GPUs per job) but in practice extended over multiple days due to cluster instability. The final RL runs for the initial Olmo 3 Think 32B spanned approximately 5 days with at least a day of training time lost due to stability issues. After the initial release of Olmo 3 , we continued our best RL run for another 21 days on 224 GPUs to produce Olmo 3.1 Think 32B.

While pretraining accounts for the majority of total GPU hours, a non-trivial share is consumed by post-training and by the repeated checkpoint evaluations required when transitioning between major training stages. These additional costs are not captured when reporting pretraining hours alone but remain significant across the model’s full development cycle. Further pretraining details, which represent the bulk of expenditure, are provided in Appendix A.2 .

## 3 Olmo 3 Base

The goal of Olmo 3 Base is to establish a strong foundation that supports a diversity of general capabilities while enabling downstream capabilities like thinking, tool-use, and instruction-following to be easily elicited during post-training. In this section, we describe our recipe for Olmo 3 Base , organized as follows:

• Modeling (Section § 3.2 ) Olmo 3 Base closely follows OLMo 2 in that it is a dense model at 7B and 32B sizes, with largely identical hyperparameters. Apart from engineering improvements that enable better training throughput, we focus on enabling a larger context window. We lay out the details in Section § 3.2 .

• Evaluation (Section § 3.3 ) To guard against overfitting Olmo 3 Base to any one capability, we greatly expand on our evaluation suite from OLMo 2 to include more benchmarks. We make small-scale experiments more reliable by systematically refining benchmark selection and usage throughout development.

• Data We introduce Dolma 3 , a collection of data to support multiple stages of base model development: ∘ \circ Pretraining (Section § 3.4 ) We train on Dolma 3 Mix , a mix of 5.9T tokens of diverse, natural data including sources like web pages, academic PDFs, code repositories, and more.

∘ \circ Midtraining (Section § 3.5 ) We train on Dolma 3 Dolmino Mix , a mix of 100B tokens combining our highest-quality pretraining data with substantial task data for math and code problems, general knowledge QA, instruction following, and more.

∘ \circ Long-context extension (Section § 3.6 ) We train on Dolma 3 Longmino Mix , a mix of 50B ( Olmo 3 Base 7B) or 100B ( Olmo 3 Base 32B) tokens combining long documents with our midtraining data.

### 3.1 Main Results for Olmo 3 Base

Tables 2 and 3 compare Olmo 3 Base 32B and 7B with leading fully-open and open-weights base models, demonstrating both the effectiveness of our evaluation design and the strong performance of Olmo 3 Base across a broad set of capabilities.

Olmo 3 Base is the best fully-open model at 32B parameters, outperforming Stanford Marin 32B and Apertus 70B. On Math and Code evaluation composites, it achieves double-digit improvements over the other fully-open 32B models and is within a few points of strong open-weight baselines. On MCQA benchmarks, its STEM and Non-STEM scores closely track Marin 32B and OLMo 2 32B and sit a few points behind the top open-weight models, while on GenQA Olmo 3 Base forms the top fully-open cluster with Marin 32B and OLMo 2 32B and is only narrowly behind Llama 3.1 70B among the open-weight baselines. At the 7B scale, Olmo 3 Base achieves the strongest Math and Code performance among fully-open models, with sizable margins over Marin 8B, Apertus 8B, and OLMo 2 7B. Compared to open-weight models, it trails only the strongest models such as Qwen and Nemotron Nano on Math and Code. In MCQA, Olmo 3 Base 7B is on par with the strongest fully-open models in both STEM and Non-STEM areas. Finally, on GenQA tasks, Olmo 3 Base outperforms all but Marin among listed fully-open models, and outperforms all but the larger Gemma 2 9B and Llama3.1 8B among listed open-weight models.

Fully-open Models Open-weight Models Olmo 3 32B Marin 32B Apertus 70B Gaperon 24B LLM 360 K2 V2 70B 3 3 3 For the K2 V2 results here, we use an updated pretraining checkpoint uploaded on Jan 22, 2026, released after Olmo 3. OLMo 2 32B Qwen 2.5 32B Gemma 3 27B Mistral 3.1 24B Seed 36B Gemma 2 27B Llama 3.1 70B OlmoBaseEval Math 61.9 49.3 39.7 20.7 72.9 53.9 64.7 63.2 59.5 15.3 57.5 62.0 GSM8k 80.6 69.1 63.0 33.3 90.9 77.6 81.1 81.3 79.3 26.9 76.3 81.2 GSM Symbolic 61.2 42.0 38.6 14.5 77.7 53.1 56.2 61.2 59.1 10.3 57.3 64.6 MATH 43.8 36.8 17.4 14.2 50.2 31.0 56.7 47.0 40.1 8.7 38.8 40.2 OlmoBaseEval Code 39.7 30.8 23.3 19.4 38.4 20.5 48.3 41.6 42.4 54.9 41.0 36.3 BigCodeBench 43.7 34.5 24.0 17.0 42.9 22.2 48.1 44.0 46.4 50.7 43.4 43.4 HumanEval 65.8 52.3 32.5 31.2 61.1 29.4 65.6 62.1 65.5 71.3 57.5 57.4 DeepSeek LeetCode 2.0 1.3 1.2 0.0 3.1 0.8 8.0 5.8 0.1 13.0 4.7 0.2 DS 1000 29.4 26.3 17.8 11.0 28.0 20.4 43.3 34.3 36.3 44.0 29.7 29.5 MBPP 59.6 52.1 37.6 36.7 55.7 37.1 69.8 60.0 61.9 72.0 61.7 55.5 MultiPL HumanEval 36.0 18.5 18.4 13.0 36.3 10.5 49.7 37.7 39.0 69.2 40.3 32.2 MultiPL MBPPP 41.5 30.5 31.3 26.5 41.5 23.2 53.6 47.2 47.7 63.8 49.7 35.9 OlmoBaseEval MC STEM \textbf{{OlmoBaseEval} \penalty\ MC}_{\textbf{\penalty\ STEM}} 74.5 75.9 70.0 56.2 75.7 75.3 82.2 80.2 81.5 83.4 75.6 80.1 ARC MC 94.7 93.4 90.7 72.7 93.5 94.4 97.0 95.8 96.2 97.3 94.1 95.2 MMLU STEM 70.8 68.4 57.8 45.3 66.5 64.7 79.7 74.9 76.1 82.8 65.8 70.0 MedMCQA MC 57.6 61.8 55.9 42.6 62.5 60.2 68.8 64.7 68.8 69.6 61.8 67.8 MedQA MC 53.8 60.8 52.4 35.4 61.1 62.2 68.4 68.7 70.4 70.1 61.0 72.3 SciQ MC 95.5 95.1 93.3 84.9 94.8 95.1 97.1 96.8 96.3 97.1 95.1 95.4 OlmoBaseEval MC Non-STEM \textbf{{OlmoBaseEval} \penalty\ MC}_{\textbf{\penalty\ Non-STEM}} 85.6 84.5 78.5 64.1 84.0 84.2 89.3 86.7 87.9 89.0 83.2 86.1 MMLU Humanities 78.3 78.9 74.1 56.7 78.4 79.7 85.0 80.5 82.7 85.7 79.3 83.4 MMLU Social Sci. 84.0 83.7 79.2 58.9 84.1 84.5 88.4 86.2 88.6 90.1 85.8 87.4 MMLU Other 75.1 75.4 70.1 55.4 77.1 75.6 81.2 80.2 81.9 82.4 76.9 79.4 CSQA MC 82.3 80.1 76.9 60.6 80.2 81.2 89.9 79.0 80.5 81.1 78.1 79.0 PiQA MC 85.6 90.5 79.0 72.0 87.5 87.7 93.3 90.3 91.0 92.5 89.0 91.5 SocialIQA MC 83.9 82.4 79.3 71.3 83.0 82.3 86.6 81.2 81.0 84.9 81.0 83.5 CoQA Gen2MC MC 96.4 93.9 87.5 67.3 92.2 94.4 96.8 95.8 94.9 96.9 94.3 95.1 DROP Gen2MC MC 87.2 71.0 56.5 48.0 67.6 68.6 86.6 84.6 86.5 90.1 66.6 70.3 Jeopardy Gen2MC MC 92.3 95.3 93.2 77.0 95.6 96.6 97.0 95.9 97.2 96.2 92.0 97.1 NaturalQs Gen2MC MC 78.0 81.0 71.9 47.5 80.5 78.6 79.9 82.0 84.6 81.4 74.5 82.4 SQuAD Gen2MC MC 98.2 97.6 95.7 90.0 97.4 97.4 97.9 97.7 97.9 98.1 97.5 97.7 OlmoBaseEval GenQA 79.8 80.3 75.0 65.3 75.6 79.1 68.5 73.5 78.0 76.0 72.9 81.6 HellaSwag RC 84.8 87.2 84.5 75.2 86.3 87.5 86.3 86.0 86.2 84.8 86.7 88.4 Winogrande RC 90.3 90.5 87.7 80.3 89.5 89.4 87.5 91.3 90.8 89.3 90.8 91.7 Lambada 75.7 76.7 74.8 58.3 75.3 77.0 76.2 77.5 79.3 76.1 76.9 79.6 Basic Skills 93.5 91.1 87.5 83.2 91.5 88.7 94.2 94.9 91.9 96.0 93.2 92.4 DROP 80.9 76.5 56.3 59.4 75.0 76.3 53.7 75.9 74.9 76.1 73.2 78.3 Jeopardy 75.3 80.5 77.2 58.9 77.6 79.1 74.0 82.1 80.3 77.4 80.7 84.0 NaturalQs 49.0 55.1 43.1 33.5 45.7 51.4 39.3 49.2 45.1 30.7 47.1 53.1 SQuAD 94.5 94.4 90.7 89.3 93.9 94.0 64.9 92.4 92.6 89.1 93.0 92.9 CoQA 74.1 70.7 72.8 49.8 45.6 68.7 40.4 12.4 61.1 64.4 14.9 73.9 OlmoBaseEval HeldOut LBPP 21.8 17.3 8.1 4.3 19.9 8.2 40.3 17.7 30.3 42.6 19.7 11.8 BBH 77.6 70.1 58.8 36.6 82.6 64.6 81.1 77.4 81.4 85.0 74.8 80.8 MMLU Pro MC 49.7 48.1 39.6 21.3 50.1 46.9 61.1 53.1 58.9 62.2 47.6 50.4 Deepmind Math 29.6 26.7 20.1 28.3 29.8 22.0 40.7 30.4 35.3 31.3 27.6 40.3