Edge0-145B brings 145 billion parameter AI to local laptops
The new open-source Edge0-145B model runs on personal hardware with 29GB peak memory, enabling complex coding and reasoning tasks without cloud dependencies.
この記事は英語版のみ利用可能です。
The release of Edge0-145B marks a significant shift in local artificial intelligence capabilities by making a 145-billion-parameter model runnable on personal laptops. Published on November 12, 2026, this open-source initiative includes the model weights, a unified inference framework, and deployment tools designed to eliminate the traditional trade-off between model size and on-device performance. This development allows developers to execute complex coding, reasoning, and agent-based workflows locally, bypassing the need for continuous cloud connectivity or expensive API fees.
What happened
For years, running large language models on personal hardware required compromising either on model capability or system resources. Smaller models often struggled with complex logical tasks, while larger models demanded server-grade infrastructure due to immense memory and bandwidth requirements. The Edge0-145B release challenges this status quo by demonstrating that a model with 145 billion parameters can operate effectively on a compatible laptop, such as a MacBook Pro, with a reported peak memory usage of just 29GB. This is accompanied by the release of Edge0-120B, offering another tier of performance for local deployment.
The project addresses four primary constraints that have historically limited on-device AI. First, it solves the hardware barrier by enabling large models to run without servers. Second, it ensures that these local models do not sacrifice capability, delivering coding and reasoning results that compete with frontier cloud models. Third, it supports agentic workflows where the assistant plans, writes code, debugs errors, and visually checks results for projects like 3D scenes or PCB designs. Finally, it removes the dependency on cloud providers, allowing data to remain on the device and ensuring core capabilities function offline.
This release provides the full stack for developers, including model weights and inference frameworks, enabling them to reproduce results, fine-tune the models, or integrate them directly into their products. The accompanying Edge0 App offers a way to test these capabilities immediately. By open-sourcing these tools, the creators aim to accelerate the adoption of high-performance local AI across various hardware configurations.
How it works
The ability to run a 145-billion-parameter model on consumer hardware relies on a co-design of the model architecture and the inference engine. The architecture uses a sparse expert approach where large matrices are shared across groups of experts. Specifically, each group of 64 experts shares two large matrices, while retaining small unique parameters for individual differences. Although the total parameter count is 145 billion, only approximately 10 billion parameters participate in computation during any single inference step. This design significantly reduces the amount of weight data that must be stored and moved during operation.
Training was conducted with local execution constraints in mind from the start, rather than treating on-device deployment as an afterthought involving post-training compression. The inference engine further optimizes performance by pipelining memory access across three tiers: GPU memory, system memory, and SSD storage. Frequently used "hot experts" stay in GPU memory, while a library of approximately 40GB of experts resides in system memory, ready to be loaded as needed. Vocabulary data, totaling around 48GB, is read on demand from the SSD. This overlapping of reading and computation avoids serial waits that typically bottleneck large model inference.
Scheduling plays a critical role by utilizing both the CPU and GPU in parallel. The GPU handles computations for cached experts, while the CPU manages the remaining workload. The system also employs speculative decoding, which verifies multiple candidate tokens simultaneously. This technique spreads the cost of reading weights across more output tokens, contributing to the reported decode speed of over 80 tokens per second. These combined strategies allow the model to sustain long contexts of 128K to 256K tokens without significant speed degradation.
Key details
- Model Size: Edge0-145B contains 145 billion parameters, with Edge0-120B also released as part of the suite.
- Memory Usage: The model requires a peak memory of 29GB on compatible laptops, a significant reduction compared to traditional expectations for models of this scale.
- Performance Speed: It achieves over 1,100 tokens per second for prompt prefill and maintains a decode speed of more than 80 tokens per second.
- Power Consumption: Typical power usage is approximately 120W, allowing sustained operation without throttling in tested setups.
- Benchmark Scores: The model scores 64.6 on DeepSWE 1.1 for coding, 90.4 on GPQA-Diamond for reasoning, and 80.6 on IFBench for instruction following.
- Context Window: Supports long-context tasks with a window of 128K to 256K tokens, maintaining consistent speed throughout extended interactions.
Why it matters
For software engineers and technical founders, this release removes several barriers to building privacy-first and offline-capable AI applications. The ability to run a model with frontier-level reasoning capabilities on a laptop means that sensitive data, such as proprietary codebases or internal documents, never needs to leave the local environment. This eliminates the legal and security complexities associated with sending data to third-party cloud providers. Additionally, the absence of recurring API fees makes it economically viable to deploy AI agents that perform extensive background tasks, such as debugging or generating complex visual assets, without worrying about token costs.
The performance metrics suggest that local AI is no longer limited to simple chatbots or basic summarization tasks. With competitive scores in coding and reasoning benchmarks, developers can rely on Edge0-145B for substantial engineering assistance. The model’s ability to plan, execute, and verify code locally opens up new possibilities for integrated development environments and autonomous coding agents. This shift empowers teams to build more robust, self-contained tools that do not depend on network stability or external service availability.
What you can do
- Download the Edge0-145B and Edge0-120B model weights from the official open-source repository.
- Install the unified inference framework to test the model’s performance on your specific hardware configuration.
- Evaluate the model’s coding capabilities by running it against your existing codebase for debugging or refactoring tasks.
- Experiment with the agentic features by setting up local workflows that involve planning and visual verification for projects like game development or hardware design.
- Fine-tune the model on your proprietary datasets to create specialized assistants that remain entirely on-premise.
- Integrate the deployment tools into your product pipeline to offer users a high-performance, offline AI experience.



