Artificial intelligence company Edge0 has released two large open-weight AI models, Edge0-145B and Edge0-120B, alongside an inference framework and deployment tools designed to enable local execution on compatible hardware.
The release, announced on 12 November 2026, aims to expand access to advanced AI capabilities for coding, reasoning, and agent-based tasks without requiring users to rely exclusively on cloud-based services.
According to Edge0, the models are designed to address challenges associated with running large AI systems locally, including memory requirements, computational demands, and the cost of accessing cloud AI services.
The company says its approach combines a sparse mixture-of-experts architecture with a memory-management system that distributes model data across graphics processing unit (GPU) memory, system memory and solid-state drives (SSDs).
Edge0 Introduces Two Large AI Models
Edge0-145B and Edge0-120B are based on a latent mixture-of-experts architecture derived from Qwen3.8-Flash-Next, according to the company’s technical documentation.
The models contain 145 billion and 120 billion parameters, respectively. However, Edge0 says only about 10.12 billion parameters in the 145B model are active during each inference step, limiting the computation required for individual tasks.
The company has released quantised model weights in Q3_K and NVFP4 formats, alongside its inference software.
Edge0 says the models are intended for complex coding, reasoning and AI agent workflows, including tasks that involve planning, executing code and using tools.
The company has also published technical documentation and model files through its official GitHub and Hugging Face repositories.
How Edge0 Enables Local AI Deployment
Running large AI models on personal computers has traditionally required substantial memory and processing resources.
Edge0 says its architecture is designed to reduce these requirements by distributing model data across different memory tiers.
Under the company’s approach, frequently used model components can remain in GPU memory, while other data is stored in system memory and retrieved from SSD storage when needed.
The system also coordinates CPU and GPU operations to support model execution.
Edge0 reports that Edge0-145B can operate with a peak memory footprint of approximately 29GB under its internal testing conditions. The company says this could make the model suitable for some computers with 48GB of available memory.
However, the reported memory figure should not be interpreted as a universal hardware requirement. Actual performance and compatibility depend on the model format, computer configuration, memory availability and inference software.
The company’s published deployment instructions for the latest 145B and 120B releases specify compatible NVIDIA GPU hardware on Linux or Windows.
Edge0 Reports Performance Across AI Benchmarks
Edge0 has published benchmark results comparing its models with several other AI systems across coding, reasoning, knowledge and instruction-following tasks.
The company reports that Edge0-145B achieved a mean score of 72.8 across six benchmarks, while Edge0-120B recorded a mean score of 68.9.
Edge0 also reports results of 86.52 on Terminal-Bench 2.1 and 64.6 on DeepSWE 1.1 for its 145B model.
These figures are based on the company’s evaluations. They have not been independently established as universal performance outcomes, and benchmark results can vary by testing environment, model configuration, and evaluation method.
The company says it is releasing its evaluation documentation and reproduction instructions to support further examination of the results.
Potential Implications for Businesses and Developers
The release could be relevant to developers and organisations exploring alternatives to cloud-based AI deployment.
Local AI systems can offer potential advantages in data control, privacy and reduced dependence on internet connectivity, depending on how they are configured and used.
For businesses handling sensitive information, locally deployed models may provide an additional option for processing certain workloads within their own computing environments.
Developers could also use local models for coding assistance, experimentation and AI agent workflows without necessarily relying on a continuously connected hosted service.
However, local deployment does not automatically guarantee data privacy or lower operating costs. Organisations must still consider hardware investment, electricity consumption, software licensing, security and maintenance requirements.
For Nigeria and other African markets, the development raises questions about access to advanced AI computing, particularly where cloud infrastructure, connectivity, and service costs remain important considerations.
The hardware requirements of Edge0’s latest models may also limit adoption among users without access to compatible GPUs and sufficient memory.
Edge0 Expands Its On-Device AI Development
The release follows Edge0’s earlier development of smaller AI models, including its Edge0-35B and Edge0-8B systems.
The company has been developing AI technologies intended for deployment across personal computers, mobile devices and other consumer hardware.
Read More
- INEC Removes 1.6 Million Invalid Voter Records, Identifies 54 Officials Across 10 States
- China’s Moonshot Drops Kimi K3: A 2.8-Trillion Parameter Shockwave Rattles Silicon Valley
- Alibaba Unveils Qwen3.8-Max, Raising the Stakes in the Global AI Race
- What Is Google TurboQuant? Understanding the AI Memory Compression Breakthrough
- Apple Plans Price Hikes as AI-Driven Chip Costs Rise
- Clarivate Unveils AI50 List of Leading Artificial Intelligence Innovators
With the introduction of Edge0-145B and Edge0-120B, the company is extending its model range towards more demanding coding and agent-based tasks.
The latest release reflects a broader industry effort to explore how advanced AI capabilities can operate outside large cloud-based infrastructure.
While Edge0’s performance and memory claims will require further evaluation across independent testing environments, its release provides developers and researchers with additional open-weight models and deployment tools to examine.
