Olmo 3 vs. Leanstral 1.5

Olmo 3 constitutes an extensive series of open models that include versions with 7 billion and 32 billion parameters, delivering outstanding performance in areas such as base functionality, reasoning, instruction, and reinforcement learning, all while ensuring transparency throughout the development process, including access to raw training datasets, intermediate checkpoints, training scripts, extended context support (with a remarkable window of 65,536 tokens), and provenance tools. The backbone of these models is derived from the Dolma 3 dataset, which encompasses about 9 trillion tokens and employs a thoughtful mixture of web content, scientific research, programming code, and comprehensive documents; this meticulous strategy of pre-training, mid-training, and long-context usage results in base models that receive further refinement through supervised fine-tuning, preference optimization, and reinforcement learning with accountable rewards, leading to the emergence of the Think and Instruct versions. Importantly, the 32 billion Think model has earned recognition as the most formidable fully open reasoning model available thus far, showcasing a performance level that closely competes with that of proprietary models in disciplines such as mathematics, programming, and complex reasoning tasks, highlighting a considerable leap forward in the realm of open model innovation. This breakthrough not only emphasizes the capabilities of open-source models but also suggests a promising future where they can effectively rival conventional closed systems across a range of sophisticated applications, potentially reshaping the landscape of artificial intelligence.

Leanstral 1.5 is a model under the Apache-2.0 license, crafted for proficient proof engineering within Lean 4, with the goal of improving both the functionality and accessibility of formal verification. This model features a remarkable total of 119 billion parameters, of which 6 billion are actively utilized, representing a major leap in efficiency for tasks related to theorem proving, agent-based proof engineering, and practical code verification. The evolution of Leanstral 1.5 included an extensive three-phase training regimen, which encompassed mid-training, supervised fine-tuning, and reinforcement learning through CISPO. In a multiturn environment, the model is designed to accept a theorem statement, propose a proof, and iteratively adjust its strategy based on input from the Lean compiler until the proof compiles successfully or resources run out. Operating in the code agent framework, Leanstral acts similarly to a developer navigating through a file system, enabling it to modify files, run bash commands, and engage with the Lean language server to track goals, errors, and type information in real time. This cutting-edge methodology not only simplifies the proof engineering workflow but also significantly enriches the overall user experience in formal verification tasks, making the process both more efficient and user-friendly. Furthermore, the integration of these capabilities positions Leanstral 1.5 as a transformative tool in the landscape of formal verification.