DeepSWE
DeepSWE is a fully open source, state-of-the-art coding agent built on top of the Qwen3-32B foundation model and trained exclusively via reinforcement learning (RL), without supervised finetuning or distillation from proprietary models. It is developed using rLLM, Agentica’s open source RL framework for language agents. DeepSWE operates as an agent; it interacts with a simulated development environment (via the R2E-Gym environment) using a suite of tools (file editor, search, shell-execution, submit/finish), enabling it to navigate codebases, edit multiple files, compile/run tests, and iteratively produce patches or complete engineering tasks. DeepSWE exhibits emergent behaviors beyond simple code generation; when presented with bugs or feature requests, the agent reasons about edge cases, seeks existing tests in the repository, proposes patches, writes extra tests for regressions, and dynamically adjusts its “thinking” effort.
Learn more
Devstral 2
Devstral 2 is a next-generation, open source agentic AI model tailored for software engineering: it doesn’t just suggest code snippets, it understands and acts across entire codebases, enabling multi-file edits, bug fixes, refactoring, dependency resolution, and context-aware code generation. The Devstral 2 family includes a large 123-billion-parameter model as well as a smaller 24-billion-parameter variant (“Devstral Small 2”), giving teams flexibility; the larger model excels in heavy-duty coding tasks requiring deep context, while the smaller one can run on more modest hardware. With a vast context window of up to 256 K tokens, Devstral 2 can reason across extensive repositories, track project history, and maintain a consistent understanding of lengthy files, an advantage for complex, real-world projects. The CLI tracks project metadata, Git statuses, and directory structure to give the model context, making “vibe-coding” more powerful.
Learn more
DeepCoder
DeepCoder is a fully open source code-reasoning and generation model released by Agentica Project in collaboration with Together AI. It is fine-tuned from DeepSeek-R1-Distilled-Qwen-14B using distributed reinforcement learning, achieving a 60.6% accuracy on LiveCodeBench (representing an 8% improvement over the base), a performance level that matches that of proprietary models such as o3-mini (2025-01-031 Low) and o1 while using only 14 billion parameters. It was trained over 2.5 weeks on 32 H100 GPUs with a curated dataset of roughly 24,000 coding problems drawn from verified sources (including TACO-Verified, PrimeIntellect SYNTHETIC-1, and LiveCodeBench submissions), each problem requiring a verifiable solution and at least five unit tests to ensure reliability for RL training. To handle long-range context, DeepCoder employs techniques such as iterative context lengthening and overlong filtering.
Learn more
MiniMax M2.5
MiniMax M2.5 is a frontier AI model engineered for real-world productivity across coding, agentic workflows, search, and office tasks. Extensively trained with reinforcement learning in hundreds of thousands of real-world environments, it achieves state-of-the-art performance in benchmarks such as SWE-Bench Verified and BrowseComp. The model demonstrates strong architectural thinking, decomposing complex problems before generating code across more than ten programming languages. M2.5 operates at high throughput speeds of up to 100 tokens per second, enabling faster completion of multi-step tasks. It is optimized for efficient reasoning, reducing token usage and execution time compared to previous versions. With dramatically lower pricing than competing frontier models, it delivers powerful performance at minimal cost. Integrated into MiniMax Agent, M2.5 supports professional-grade office workflows, financial modeling, and autonomous task execution.
Learn more
