Nous Research on Jan. 7, 2026 released NousCoder-14B, an open-source coding model it trained in four days on 48 Nvidia B200 GPUs and reports achieves 67.87% accuracy on LiveCodeBench v6—7.08 percentage points better than the Alibaba Qwen3-14B base model.
- Open-source NousCoder-14B released by Nous Research, backed by Paradigm
- 67.87% on LiveCodeBench v6; +7.08 percentage points vs Qwen3-14B base
- Trained in four days using 48 Nvidia B200 GPUs
What happened
Nous Research published NousCoder-14B on Jan. 7, 2026. The startup says the model was trained in four days on 48 Nvidia B200 GPUs and that a technical report accompanying the release shows a 67.87 percent accuracy on LiveCodeBench v6, a benchmark of competitive programming problems from August 2024 to May 2025.
According to Nous Research, that LiveCodeBench result represents a 7.08 percentage point improvement over the base model used in training, Alibaba's Qwen3-14B. The company is positioning NousCoder-14B as an open-source alternative in the crowded coding-assistant market.
Why it matters
The release highlights intensifying competition between open-source efforts and proprietary coding agents: Anthropic's Claude Code has dominated developer conversation recently, and NousCoder-14B shows an open-source project can post competitive benchmark results quickly using modern GPU clusters.
For developers and teams, a viable open-source coding model can lower barriers to experimentation, integration, and customization compared with closed systems, while benchmarks like LiveCodeBench provide a common yardstick for comparison even as real-world utility depends on robustness, toolchain integration, and developer feedback.
What to watch next
Look for independent evaluations, community adoption, and real-world developer reports that confirm whether NousCoder-14B's benchmark gains translate to practical coding workflows and complex tasks beyond competitive programming.
Also watch how Nous Research and rivals iterate: updates to model capabilities, fine-tuning options, safety and reliability improvements, and comparisons with Anthropic's Claude Code will shape which tools gain traction among engineers and enterprise buyers.