Apriel-1.5-15b-Thinker

1 Introduction

• Organizations with privacy or compliance needs face constraints on on-premises deployment, while training and inference costs can determine production-scale feasibility.The paper frames this as whether an economical compact, open multimodal model can achieve frontier-level reasoning.Source: §1 · p. 2
• The training approach omits reinforcement learning and preference optimization, allowing a clearer assessment of the mid-training recipe under computational constraints.The paper says the release focuses on maximizing the base model's potential through mid-training.Source: §1 · pp. 2–3
• 52 on the Artificial Analysis Intelligence Index matches DeepSeek-R1-0528, while performance across ten multimodal benchmarks averages only 5 points behind Gemini-2.5-Flash and Claude Sonnet-3.7.The model is described as operating within single-GPU deployment constraints.Source: §1 · p. 3
• Reported evaluations span text and multimodal tasks, including AIME’25 at 88%, MMMU at 70.2%, and CharXiv at 88.2%.The contribution summary also names IFBench, MathVista, and τ 2-Bench Telecom results.Source: §1 · p. 3
• The work releases model weights, training recipes, and evaluation artifacts under a permissive license to enable reproducibility and further study.The paper describes the model as an open-source compact multimodal foundation model.Source: §1 · p. 3

2 Architecture and Model Upscaling

• Pixtral-12B-Base-2409 provides the base multimodal architecture, with a vision encoder connected to a decoder through a two-layer projection network.The base follows the LLaVA architecture.Source: §2 · p. 3
• Depth upscaling increases the decoder from 40 to 48 hidden layers and trains it on replay tokens plus text from diverse domains.The text corpus includes web content, technical literature, mathematics, programming, and StackExchange; the method is designed to retain single-high-end-GPU deployability.Source: §2 · pp. 3–4
• Projection realignment trains on captioning, multimodal instruction-response, and document-understanding data while freezing the vision encoder and decoder.This updates the projection network between the upscaling stage and subsequent training.Source: §2 · p. 4
• Both stages use sequence length 8192 and learning rate 5e-5; six intermediate upscaling checkpoints are averaged before projection realignment.The final projection-realignment checkpoint is used for subsequent training.Source: §2 · p. 4

3 Continual Pretraining (CPT)

• The CPT curriculum has two stages: foundational text reasoning and image understanding first, followed by further visual-reasoning training.The first stage strengthens textual reasoning and image understanding; the second targets visual reasoning.Source: §3 · p. 4

3.1 CPT Stage 1

• Stage 1 trains on 50% text-only reasoning and knowledge tokens, 20% replay tokens, and 30% multimodal data for document, chart, image, OCR, and visual reasoning tasks.The vision encoder, projection network, and decoder remain unfrozen during this stage.Source: §3.1 · p. 4

3.2 CPT Stage 2

• Stage 2 converts raw images into synthetic task-focused samples targeting spatial structure, compositionality, and fine-grained perception.The tasks include image reconstruction, visual matching, object detection, and counting; task-dependent difficulty modulation supports robust grounding and reasoning. The vision encoder is frozen while the projection network and decoder are updated.Source: §3.2 · pp. 4–5
• +9.65 points on MathVerse Vision Dominant, +5.98 points on CharXiv Descriptive, and +3.7 points on AI2D Test are reported for Stage 2 over Stage 1.The comparison uses two small-scale SFT experiments initialized from the respective CPT checkpoints and 17k text-based reasoning samples; Table 1 covers multiple vision benchmarks.Source: §3.2 · p. 5, Table 1 · p. 5

4 Supervised Fine Tuning (SFT)

• The authors select gpt-oss-120b as annotator for compute efficiency after small-scale ablations found minimal performance differences from DeepSeek-R1-0528 annotations.The table caption reports similar overall benchmark performance for the two annotator models.Source: §4 · p. 5, Table 2 · p. 6
• The data pipeline filters duplicates, unsafe or low-quality samples, and incorrect responses, using LLM judging and execution-based verification where applicable.It also checks expected output formats, applies a consistent chat template, and removes benchmark overlaps.Source: §4 · pp. 5–6
• Millions of instruction-response pairs span mathematics, coding, science, tool calling, general reasoning, knowledge-intensive tasks, and other domains, with explicit reasoning steps followed by final answers.The dataset also includes conversations, instruction-following, security, content moderation, and robustness samples.Source: §4 · p. 6
• After four epochs of large-scale SFT, two smaller runs use a stratified 25% subset and longer sequences; averaging their weights produces the final checkpoint.Only the decoder is updated, and loss is computed on responses; the authors report inexpensive gains in overall and long-context performance from the smaller runs.Source: §4 · p. 6

5.1 Text evaluation

• Text results use the Artificial Analysis Intelligence Index, an independent combination metric intended to support unbiased comparisons across organizations.The authors contrast this external measure with in-house test sets and proprietary metrics.Source: §5.1 · p. 6
• The Artificial Analysis Intelligence Index combines ten heterogeneous benchmarks into a normalized measure of general intelligence across domains and evaluation difficulty.The listed benchmarks cover knowledge and reasoning, coding, mathematics, instruction following, long-context reasoning, terminal tool use, and telecom tasks.Source: §5.1 · pp. 6–7

5.2 Vision evaluation

• Vision evaluation uses VLMEvalKit to standardize data loading, prompting, post-processing, and scoring across image benchmarks.The authors state that their training was conducted primarily with images.Source: §5.2 · p. 7
• The vision suite covers multimodal reasoning, visual logic, mathematical vision, document and diagram understanding, and open-domain vision-language reasoning.Its named benchmarks include MMMU, MMMU-Pro, MMStar, LogicVista, MathVision, MathVista, MathVerse, CharXiv, AI2D, and BLINK.Source: §5.2 · p. 7
• Vision evaluations follow official or community-standard VLMEvalKit protocols with consistent prompts and inference settings across models.The suite also includes BLINK for open-domain vision-language reasoning and visual perception tasks.Source: §5.2 · p. 7

6.1 Text Benchmarks

• 87% on AIME2025, 62% on IF-Bench, and 68% on τ 2Bench (Telecom) are reported as strong results in mathematics, instruction following, and domain-specific problem solving.The authors say these results consistently outperform significantly larger open-source baselines.Source: §6.1 · p. 7
• 10% on TerminalBench-Hard is competitive with GPT-4.1, Gemini 2.5 Flash, and Qwen3-250B at 13%, and exceeds gpt-oss-20b at 6%.The benchmark covers terminal-based technical tasks.Source: §6.1 · pp. 7–8
• 52 on the Artificial Analysis Intelligence Index places Apriel-1.5-15B-Thinker among the leading small open-source models, with performance comparable to DeepSeek-R1-0528 and Gemini-2.5-Flash.Figure 2 reports that it ranks first among SOTA small open-source models; the text reports a score of 52 and the comparisons.Source: §6.1 · p. 7, Figure 2 · p. 8
• Apriel’s Table 3 scores are self-reported, and evaluation-condition differences are cited to explain divergences from Artificial Analysis results.The stated differences include judging models for AA-LCR and τ 2Bench, no language-model equality checker for AIME2025, and a shorter TerminalBench timeout.Source: §6.1 · p. 8
• The scale comparison places Apriel in Figure 3’s “most attractive quadrant,” where moderate model scale coincides with disproportionately high performance.The authors associate this placement with reduced compute requirements and faster inference while maintaining robust general capabilities.Source: §6.1 · p. 8, Figure 3 · p. 9

6.2 Vision Benchmarks

• Among image benchmarks, Apriel beats most similarly sized and some larger open-weight models, tracks 400B-parameter Llama 4 Maverick[4], and exceeds GPT-5 Minimal and Grok 4 Fast in overall average score.Figure 4 aggregates scores across ten image benchmarks, with higher average performance better.Source: Figure 4 · p. 10, §6.2 · p. 10
• 88.20% on CharXiv descriptive tasks is the third-highest score after Claude and GPT-5 (High), while 76.40% on MathVerse (Text Dominant) exceeds Claude Sonnet, Magistral, and Llama 4 Maverick[4].The authors also report strong document-centric and diagram understanding, competitive MMStar performance, and visual mathematical skills on MathVista.Source: §6.2 · pp. 10–11
• 70.22% on MMMU contrasts with 48.21% on MMMU-PRO (Vision); CharXiv descriptive performance exceeds CharXiv reasoning by 38.1 percentage points.The authors describe stronger results on visual-input tasks with substantial textual reasoning and room to improve on purely visual and demanding STEM-centric tasks.Source: §6.2 · p. 11

• A 15-billion-parameter model reaches frontier-level reasoning through data quality and structured mid-training—CPT followed by high-signal SFT—without reinforcement learning or preference optimization.Source: §7 · p. 11
• +9.65 on MathVerse Vision-Dominant during CPT accompanies strong AIME and GPQA text-reasoning results and competitive multimodal benchmarks.Source: §7 · p. 11
• Single-GPU operation delivers a favorable performance–efficiency trade-off for organizations with limited computational infrastructure.Source: §7 · p. 11
• The work focused primarily on text reasoning; future work targets more comprehensive multimodal capabilities and stronger agentic abilities for interactive workflows.Source: §7 · p. 11

Core Contributors

• The core contributor group includes 14 named contributors.The names are Shruthan Radhakrishna, Aman Tiwari, Aanjaneya Shukla, Masoud Hashemi, Rishabh Maheshwary, Shiva Krishna Reddy Malay, Jash Mehta, Pulkit Pattnaik, Saloni Mittal, Khalil Slimi, Kelechi Ogueji, Akintunde Oladipo, Soham Parikh, and Oluwanifemi Bamgbose.Source: Core Contributors · p. 13

Secondary Contributors

• Five secondary contributors are acknowledged.They are Toby Liang, Ahmed Masry, Khyati Mahajan, Sai Rajeswar Mudumba, and Vikas Yadav.Source: Secondary Contributors · p. 13

Leads & Management

• The acknowledgments recognize reused upstream project work and leadership of evaluations, benchmarking, and data infrastructure.Upstream contributors named are Gopal Sarda, Anil Turkkan, Shashank Maiya, Dhruv Jhamb, Jishnu S Nair, Akshay Kalkunte, Bidyapati Pradhan, Surajit Dasgupta, Jaykumar Kasundra, Anjaneya Praharaj, Sourabh Surana, Lakshmi Sirisha Chodisetty, Abhigya Verma, Abhishek Bharadwaj, Nandhakumar Kandasamy, Naman Gupta, and Segan Subramanian. Anil Madamala led evaluations and benchmarking; Segan Subramanian and Vipul Mittal led data infrastructure.Source: Leads & Management · p. 13