Edge0 Open-Sources 145B-Parameter AI Model for Local Laptop Execution
Edge0 open-sourced its 145 billion parameter Edge0-145B AI model on November 12, 2026, enabling frontier-class coding and agent workloads to run locally on laptops. This release, which includes the Edge0-120B inference framework and deployment tools, supports macOS, Windows, and Linux, including Apple M-series chips and NVIDIA platforms. For broader context, explore our AI Tools Pricing.
Edge0-145B: Architecture and Performance
The Edge0-145B model utilizes a sparse mixture-of-experts (MoE) architecture. This design means that while the model has 145 billion total parameters, only approximately 10 billion parameters are active during each inference step. This approach contributes to its ability to run on consumer hardware.
According to company-reported benchmarks, Edge0-145B demonstrates notable performance on compatible laptops. It achieves a peak memory usage of 29GB, with decoding speeds reported above 80 tokens per second and prefill speeds exceeding 1,100 tokens per second. The typical power consumption is around 120W, and the model supports a context window ranging from 128K to 256K tokens. Edge0 states that the model was trained specifically for its on-device architecture from the outset, rather than being a compressed version of a cloud-based model.
Benchmark Comparisons and Agentic Capabilities
Edge0's internal testing indicates that Edge0-145B outperforms several compact frontier cloud models across various benchmarks. On DeepSWE 1.1, it scored 64.6, surpassing Qwen3.8-Flash-Next (58.7) and DeepSeek-V4-Flash-0731 (54.4). For Terminal-Bench 2.1, Edge0-145B achieved 86.52, compared to Claude's 78.2. On IFBench, it scored 80.6, exceeding Claude-Opus-4.6 (Max) at 62.5. On GPQA-Diamond, it scored 90.4, slightly below DeepSeek-V4-Flash-0731 (90.8) and Claude-Opus-4.6 (Max) (91.3). All these figures are based on Edge0's internal evaluations.
A key feature of Edge0-145B is its support for agentic workflows that operate entirely offline. Demonstrations include the model generating flight-control and visual-recognition code for a drone from a single instruction, enabling it to take off, navigate obstacles, and locate a red backpack indoors. Other reported demonstrations include creating a Manhattan-scale 3D scene in Blender within an hour, developing playable games, producing a one-hour explainer video, and routing a PCB in KiCad with an assembly animation rendered in Unreal Engine 5.
Availability and Platform Support
The Edge0-145B and Edge0-120B models, along with the unified inference framework, are available for download. The weights are hosted on Hugging Face, and the framework is accessible on GitHub. The Edge0 App is available as a v0.2.0 release.
Day-one support for this release includes macOS, Windows, and Linux operating systems. This encompasses compatibility with Apple M-series chips and NVIDIA platforms, such as RTX Spark, RTX 4090, and RTX 5090.
Implications for Local AI Development
The release of a 145 billion parameter open-weights model capable of running on a laptop represents a notable development for developers and enterprises. Historically, on-device models were limited to roughly 10 billion parameters due to hardware memory constraints; for example, Apple Intelligence operates at approximately 3B parameters, and Google's Gemini Nano ships in 1.8B and 3.25B variants. This new capability changes the landscape by offering several advantages:
- Cost Efficiency: Local inference eliminates cloud token or API fees, which can lead to significant cost savings as usage scales.
- Data Privacy: Sensitive code and documents remain on the local machine, enhancing data security and privacy.
- Offline Functionality: Core AI capabilities continue to function without an internet connection, ensuring uninterrupted workflow.
Edge0's previous 35B phone model achieved the number one position on Hugging Face's Trending chart and has accumulated over 15 million downloads, indicating an established demand for local AI solutions.
Conclusion
Edge0's open-sourcing of its 145B-parameter model and associated tools provides a new option for developers seeking to deploy advanced AI capabilities directly on consumer hardware. This development could accelerate the adoption of local AI for complex tasks, particularly in coding and agentic workflows, by addressing concerns related to cost, privacy, and offline access. The availability across major operating systems and hardware platforms suggests a broad potential impact on how AI models are utilized outside of cloud environments.
Sources
About the Author
Albert Schaper is a co-founder of Best-AI.org. He focuses on product strategy, AI adoption, practical tool selection, and educational content that helps users compare AI products with clearer context.
More from AlbertWas this article helpful?
Found outdated info or have suggestions? Send us a note.