Nous Research Releases Open-Source Competitive Programming Model NousCoder-14B

Nous Research released open model weights and source code for NousCoder-14B on August 8, 2026. Built as a reinforcement-learned fine-tune of Alibaba’s Qwen3-14B baseline, the 14-billion-parameter model is optimized specifically for Olympiad-level competitive programming and complex algorithmic problem-solving. The release includes open-weight artifacts on Hugging Face alongside the full reinforcement learning training harness and environment code on GitHub under the Apache 2.0 license.

Reinforcement Learning Architecture and Training Pipeline

NousCoder-14B was post-trained using reinforcement learning with verifiable rewards, building upon the Atropos RL framework developed by Nous Research. Training was completed over four days using a dedicated cluster of 48 Nvidia B200 GPUs.

Rather than relying on static text matching or manual feedback, the post-training process utilizes execution-based feedback from actual code rollouts:

• Verifiable Reward Signals: The model receives scalar rewards based on execution results against test cases. Generating code that passes all hidden input and output test cases yields a reward of +1. Failing any test case, returning incorrect outputs, or violating performance constraints results in a reward of -1.
• Sandboxed Parallel Verification: To execute generated solutions safely without resource contention on training GPUs, Nous Research integrated Modal autoscaling containers. Each rollout executes within an isolated container that evaluates the program against hundreds of test cases per problem.
• Execution Constraints: Verification containers enforce a strict 15-second runtime limit and a 4 GB memory ceiling per rollout to reflect standard competitive programming contest rules.

The system uses asynchronous execution pipelines to overlap inference rollouts with verification tasks, maintaining high GPU utilization throughout the training run.

Training Dataset and Benchmark Performance

The dataset and evaluation protocol focus entirely on competitive programming challenges where solution correctness can be objectively verified through program execution.

The training corpus comprises 24,000 competitive programming problems drawn from three sources:

• TACO Verified: Curated algorithmic problems paired with verified test suites.
• PrimeIntellect SYNTHETIC-1: Synthetic code generation tasks designed for automated verification.
• LiveCodeBench (Pre-August 2024): Historical contest problems published prior to July 31, 2024.

For evaluation, Nous Research tested NousCoder-14B against the un-contaminated LiveCodeBench v6 benchmark, which contains 454 competitive programming problems published between August 1, 2024, and May 1, 2025. On this test set, NousCoder-14B achieved a Pass@1 accuracy of 67.87%, marking a 7.08 percentage point improvement over the Qwen3-14B base model score of 60.79%.

Context Extension for Hard Algorithmic Reasoning

To support extended reasoning traces and complex code generations required for hard contest problems, the training pipeline implemented iterative context window expansion:

1. Initial Training Phase: The model began RL post-training configured with a 32,768-token context window.
2. Mid-Training Extension: At training step 50, the context window was expanded to 40,960 tokens to accommodate longer solution attempts and detailed mathematical step-by-step reasoning.
3. Evaluation Scaling: The evaluation configuration verified context handling up to 81,920 tokens, allowing the model to maintain coherence across lengthy algorithmic proofs and edge-case validations.

The extended token capacity prevents truncation during multi-step logic deductions, ensuring that complex algorithms with extensive input specifications retain full context during code generation.

Local Hardware Execution and Deployment Setup

Occupying a 14-billion parameter footprint, NousCoder-14B offers specialized algorithmic performance while remaining deployable on local developer hardware without relying on third-party cloud APIs. Depending on precision settings—such as 16-bit Brain Floating Point (BF16) or 8-bit/4-bit quantization—the model requires between 9 and 30 gigabytes of video memory.

Engineers can serve NousCoder-14B locally using open-source inference engines:

• vLLM: High-throughput serving framework featuring PagedAttention and optimized tensor parallelism.
• Ollama: Lightweight application interface for local CPU and GPU model execution.
• llama.cpp: C/C++ backend designed for quantized GGUF model execution on consumer workstations and Apple Silicon.

To serve NousCoder-14B locally using vLLM, developers can initialize an API endpoint from the terminal:

vllm serve NousResearch/NousCoder-14B \
  --tensor-parallel-size 1 \
  --max-model-len 40960 \
  --dtype bfloat16

For quantized desktop execution via Ollama, the model can be fetched and run directly:

ollama run nouscoder-14b

The release includes downloadable model weights, tokenizers, training configurations, and benchmark logs accessible across Linux and Windows developer environments under the Apache 2.0 license.