Together Ai Released Deepcoder-14b-preview: A Fully Open-source Code Reasoning Model That Rivals O3-mini With Just 14b Parameters

The request for intelligent codification procreation and automated programming solutions has intensified, fueled by a accelerated emergence successful package complexity and developer productivity needs. While earthy connection processing and wide reasoning models person surged pinch important breakthroughs, nan coding domain has knowledgeable slower progress. This lag is chiefly attributed to nan scarcity of high-quality, verifiable datasets captious for efficaciously training RL-based systems. Unlike mathematical problems, which use from a wealthiness of structured, verifiable examples online, coding tasks often suffer from noise, insufficient trial coverage, and unverifiable outputs. Consequently, advancing LLMs for codification procreation has remained a formidable situation until now.

DeepCoder-14B-Preview was released by Together AI successful collaboration pinch nan Agentica team. This powerful exemplary was fine-tuned from DeepSeek-R1-Distilled-Qwen-14B utilizing distributed reinforcement learning, and it demonstrates important advancement successful codification reasoning. With a capacity of 60.6% Pass@1 accuracy connected nan LiveCodeBench (LCB), DeepCoder-14B-Preview not only closes nan spread pinch starring models for illustration o3-mini-2025 but matches their output, each while utilizing conscionable 14 cardinal parameters, a notable feat successful ratio and capability.

The merchandise is particularly important considering nan benchmarks. DeepSeek-R1-Distill-Qwen-14B scores 53.0% connected LCB, and DeepCoder-14B-Preview demonstrates an 8% leap successful accuracy compared to its guidelines model. Also, it competes toe-to-toe pinch established models, specified arsenic o3-mini (60.9%) and o1-2024-12-17 (59.5%) successful accuracy and coding prowess. Regarding competitory coding metrics, it reaches a Codeforces standing of 1936 and a percentile of 95.3%, which are clear indicators of its real-world coding competence.

The exemplary was trained complete 2.5 weeks connected 32 H100 GPUs utilizing a curated dataset of 24,000 verifiable coding problems. This dataset was built by rigorously filtering existing resources to guarantee value and diversity. It combines problems from nan TACO Verified set, PrimeIntellect’s SYNTHETIC-1, and entries from LiveCodeBench submitted betwixt May 2023 and July 2024. The action process emphasized programmatic verification of trial cases, a minimum of 5 portion tests per problem, and deduplication to debar information contamination. This helped support training integrity and maximize RL effectiveness.

To facilitate this level of validation, DeepCoder’s training incorporated a scalable codification sandbox situation tin of executing monolithic parallel evaluations. Over 1,000 coding problems were assessed astatine each RL measurement utilizing 2 robust sandboxes, nan Together Code Interpreter and a section sandbox. These environments ensured that each model-generated solution was rigorously tested crossed aggregate portion tests, filtering retired reward hacking and encouraging genuine reasoning complete memorization.

Also, nan strategy architecture supporting DeepCoder was optimized done “verl-pipe,” an upgraded hold to nan post-training RL pipeline that doubled training velocity done systems-level improvements. This enhancement accelerates improvement cycles and provides a modular model for others looking to build aliases iterate connected akin LLMs successful open-source ecosystems.

Some Key Takeaways from nan merchandise of DeepCoder-14B-Preview include:

• DeepCoder-14B-Preview achieves 60.6% Pass@1 accuracy connected LiveCodeBench—matching o3-mini’s capacity pinch less parameters.  
• The model’s training leveraged 24K verifiable coding problems, cautiously curated to debar sound and reward hacking.  
• It was trained connected 32 H100 GPUs for 2.5 weeks, emphasizing reproducibility and strategy efficiency.  
• A dual-sandbox situation ensured meticulous and scalable codification verification during training.  
• System optimization via verl-pipe doubled training velocity and provides a reusable pipeline for early models.  
• DeepCoder is afloat open-sourced, including datasets, code, and training logs, paving nan measurement for community-driven development.  

Check out the Technical details, Model connected Hugging Face and GitHub Page. All in installments for this investigation goes to nan researchers of this project. Also, feel free to travel america on Twitter and don’t hide to subordinate our 85k+ ML SubReddit.

🔥 [Register Now] miniCON Virtual Conference connected OPEN SOURCE AI: FREE REGISTRATION + Certificate of Attendance + 3 Hour Short Event (April 12, 9 am- 12 p.m. PST) + Hands connected Workshop [Sponsored]