Cosine uses Nebius AI Cloud to train specialized coding models. Nebius describes synthetic data generation, GPU environments, same-day experimentation and scaling from smaller fine-tuning jobs to multi-node reinforcement learning runs. Cosine builds sovereign coding agents for regulated industries, critical infrastructure and complex legacy codebases, with models that can operate in fully air-gapped environments and adapt to niche programming languages, proprietary workflows and poorly documented software.
A December 2025 Nebius customer spotlight describes Cosine using Nebius Token Factory for large-scale post-training of Llama 3.3 70B and GPT-OSS-120B. The workflows combined supervised fine-tuning for specific coding environments with reinforcement learning from code execution. Nebius reports that Cosine deployed the resulting agent to one of Europe’s largest financial institutions, achieving software engineering performance equivalent to OpenAI’s o3 model within a secure, private environment.
Nebius Post-training by Nebius Token Factory: The missing layer between a minimum viable product and production
9 Dec 2025 · Nebius · Web page
Nebius describes Cosine using Nebius Token Factory for large-scale post-training of Llama 3.3 70B and GPT-OSS-120B, combining supervised fine-tuning with reinforcement learning from code execution. Nebius reports deployment of the resulting coding agent to one of Europe’s largest financial institutions, with software engineering performance equivalent to OpenAI’s o3 model in a secure, private environment.
Nebius Post-training by Nebius Token Factory: The missing layer between a minimum viable product and production9 Dec 2025 · Nebius · Web pageNebius describes Cosine using Nebius Token Factory for large-scale post-training of Llama 3.3 70B and GPT-OSS-120B, combining supervised fine-tuning with reinforcement learning from code execution. Nebius reports deployment of the resulting coding agent to one of Europe’s largest financial institutions, with software engineering performance equivalent to OpenAI’s o3 model in a secure, private environment.
Nebius How Cosine Builds AI Coding Agents for Air-Gapped Systems
4 Aug 2026 · Nebius · Video
0:14 and 1:00: Cosine co-founder says Nebius supports its experiment pace and that it began by spinning up instances, then scaled to larger and more complex experiments. The full interview describes training models for air-gapped deployments, synthetic data and reinforcement-learning work.
Nebius How Cosine Builds AI Coding Agents for Air-Gapped Systems4 Aug 2026 · Nebius · Video · Starts at 0:140:14 and 1:00: Cosine co-founder says Nebius supports its experiment pace and that it began by spinning up instances, then scaled to larger and more complex experiments. The full interview describes training models for air-gapped deployments, synthetic data and reinforcement-learning work.
Nebius customer-story post
22 Jul 2026 · LinkedIn · Nebius customer-story post · Web page
Nebius states that Cosine uses Nebius AI Cloud to generate data, launch GPU environments and scale coding-model training from smaller fine-tuning jobs to multi-node reinforcement learning runs.
Nebius customer-story post22 Jul 2026 · LinkedIn · Nebius customer-story post · Web pageNebius states that Cosine uses Nebius AI Cloud to generate data, launch GPU environments and scale coding-model training from smaller fine-tuning jobs to multi-node reinforcement learning runs.
Cosine develops AI models and autonomous coding agents for software engineers and organisations in finance, healthcare, defence and government. Its tools help teams understand existing codebases, implement changes, test software and review results through command-line, desktop and cloud interfaces. Its Lumen product family includes Outpost for coding and cybersecurity work, Scout for codebase mapping and retrieval, and the forthcoming Lumen Sovereign model for controlled environments. Enterprise deployments support on-premises and air-gapped infrastructure, allowing customers to keep code and data within their own systems.
Cosine partners with Shift on financial-sector operational resilience
Shift is integrating Cosine’s Lumen Outpost into its operations and will help test and shape the forthcoming Lumen Sovereign model. The partnership combines Shift’s resilience platform with Cosine’s models for scenario testing and analysis in regulated financial institutions.
Cosine partners with Welsh Government on Welsh-language AI
The Welsh Government will provide data, resources and expertise to improve Welsh-language capabilities in Cosine’s Lumen model. The collaboration will support representation of Welsh and bilingual usage in AI.
Kainos will contribute government and regulated-sector delivery expertise to the development of Lumen Sovereign. Its input will cover governance, assurance, security, deployment and adoption requirements.
Cosine uses Nebius AI Cloud to train specialized coding models. Nebius describes synthetic data generation, GPU environments, same-day experimentation and scaling from smaller fine-tuning jobs to multi-node reinforcement learning runs. Cosine builds sovereign coding agents for regulated industries, critical infrastructure and complex legacy codebases, with models that can operate in fully air-gapped environments and adapt to niche programming languages, proprietary workflows and poorly documented software.
A December 2025 Nebius customer spotlight describes Cosine using Nebius Token Factory for large-scale post-training of Llama 3.3 70B and GPT-OSS-120B. The workflows combined supervised fine-tuning for specific coding environments with reinforcement learning from code execution. Nebius reports that Cosine deployed the resulting agent to one of Europe’s largest financial institutions, achieving software engineering performance equivalent to OpenAI’s o3 model within a secure, private environment.
Nebius Post-training by Nebius Token Factory: The missing layer between a minimum viable product and production
9 Dec 2025 · Nebius · Web page
Nebius describes Cosine using Nebius Token Factory for large-scale post-training of Llama 3.3 70B and GPT-OSS-120B, combining supervised fine-tuning with reinforcement learning from code execution. Nebius reports deployment of the resulting coding agent to one of Europe’s largest financial institutions, with software engineering performance equivalent to OpenAI’s o3 model in a secure, private environment.
Nebius Post-training by Nebius Token Factory: The missing layer between a minimum viable product and production9 Dec 2025 · Nebius · Web pageNebius describes Cosine using Nebius Token Factory for large-scale post-training of Llama 3.3 70B and GPT-OSS-120B, combining supervised fine-tuning with reinforcement learning from code execution. Nebius reports deployment of the resulting coding agent to one of Europe’s largest financial institutions, with software engineering performance equivalent to OpenAI’s o3 model in a secure, private environment.
Nebius How Cosine Builds AI Coding Agents for Air-Gapped Systems
4 Aug 2026 · Nebius · Video
0:14 and 1:00: Cosine co-founder says Nebius supports its experiment pace and that it began by spinning up instances, then scaled to larger and more complex experiments. The full interview describes training models for air-gapped deployments, synthetic data and reinforcement-learning work.
Nebius How Cosine Builds AI Coding Agents for Air-Gapped Systems4 Aug 2026 · Nebius · Video · Starts at 0:140:14 and 1:00: Cosine co-founder says Nebius supports its experiment pace and that it began by spinning up instances, then scaled to larger and more complex experiments. The full interview describes training models for air-gapped deployments, synthetic data and reinforcement-learning work.
Nebius customer-story post
22 Jul 2026 · LinkedIn · Nebius customer-story post · Web page
Nebius states that Cosine uses Nebius AI Cloud to generate data, launch GPU environments and scale coding-model training from smaller fine-tuning jobs to multi-node reinforcement learning runs.
Nebius customer-story post22 Jul 2026 · LinkedIn · Nebius customer-story post · Web pageNebius states that Cosine uses Nebius AI Cloud to generate data, launch GPU environments and scale coding-model training from smaller fine-tuning jobs to multi-node reinforcement learning runs.
Cosine develops AI models and autonomous coding agents for software engineers and organisations in finance, healthcare, defence and government. Its tools help teams understand existing codebases, implement changes, test software and review results through command-line, desktop and cloud interfaces. Its Lumen product family includes Outpost for coding and cybersecurity work, Scout for codebase mapping and retrieval, and the forthcoming Lumen Sovereign model for controlled environments. Enterprise deployments support on-premises and air-gapped infrastructure, allowing customers to keep code and data within their own systems.
Cosine partners with Shift on financial-sector operational resilience
Shift is integrating Cosine’s Lumen Outpost into its operations and will help test and shape the forthcoming Lumen Sovereign model. The partnership combines Shift’s resilience platform with Cosine’s models for scenario testing and analysis in regulated financial institutions.
Cosine partners with Welsh Government on Welsh-language AI
The Welsh Government will provide data, resources and expertise to improve Welsh-language capabilities in Cosine’s Lumen model. The collaboration will support representation of Welsh and bilingual usage in AI.
Kainos will contribute government and regulated-sector delivery expertise to the development of Lumen Sovereign. Its input will cover governance, assurance, security, deployment and adoption requirements.
