Together AI drops a 14B coding model rivaling o3-mini

DeepCoder-14B-Preview is a newly open-sourced code reasoning model released jointly by Together AI and the Agentica team. It delivers reasoning performance on par with OpenAI's o3-mini while running on just 14B parameters.

Built on top of Deepseek-R1-Distilled-Qwen-14B using distributed reinforcement learning, the model hits a 60.6% Pass@1 accuracy on LiveCodeBench. That marks an 8% jump over its base model and matches both o3-mini and o1.

How it was built

While math reasoning models advanced quickly with reinforcement learning, coding lagged behind due to a lack of clean, verifiable reward signals. Existing coding datasets are often too trivial or lack adequate unit tests, which tempts models to exploit reward hacking by memorizing common outputs.

To fix this, the team curated a pristine training set of 24K verifiable coding problems. The mix consists of 7.5K problems from TACO Verified, 16K from PrimeIntellect's SYNTHETIC-1, and 600 from LiveCodeBench.

Quality checks were strict. Every problem had to pass programmatic verification with an official solution and include at least five distinct unit tests. The team also conducted thorough deduplication to guarantee zero contamination with test benchmarks.

Training took 2.5 weeks across 32 Nvidia H100 GPUs. To handle running code tests on thousands of candidate solutions without stalling the cluster, the team built and released verl-pipe, an extension that accelerates end-to-end RL training by 2x using parallel sandboxes.

What is open and how to run it

This is a complete open-source release. Together AI and Agentica have published the model weights alongside the entire 24K problem dataset, training scripts, execution logs, and system optimizations.

Everything is available in their public GitHub repository. Developers looking to reproduce the run or fine-tune their own models can plug in the provided verl-pipe framework and run evaluations through either the beta Together Code Interpreter or their local sandbox.

Why it matters

For engineering teams building developer tooling or running local coding agents, DeepCoder-14B is a game changer. It proves that you do not need massive proprietary models or closed APIs to handle difficult programming logic.

We think the real breakthrough is the open recipe. By open-sourcing the data filtering pipeline and the 2x faster RL framework, the creators have given everyone a practical roadmap to train top-tier coding models on a modest compute budget.