Sources checked
What is the Edge0-145B?
Edge0 released Edge0-145B, a 145-billion-parameter open-source model designed to run on personal laptops. The company calls it the first open-source on-device model at this scale, bringing coding, reasoning, and agent capabilities that rival frontier cloud models to consumer hardware.
Alongside Edge0-145B, the company released Edge0-120B, a unified inference framework, and deployment tools.
Both model weights are available on Hugging Face, and the full stack is open source so developers can reproduce the results, evaluate or fine-tune the models, and integrate them into their own products.
The model uses a mixture-of-experts architecture where only about 10 billion parameters are active on each inference step, despite the 145 billion total.
This design keeps memory and compute demands within consumer hardware limits while preserving the capability of a much larger model.
Edge0 trained the model in its on-device architecture from the outset, avoiding treating local deployment as a compression or pruning step after training a cloud model.
In the company's laptop tests, Edge0-145B reached 29GB peak memory, processed prompts above 1,100 tokens per second, generated at over 80 tokens per second, and drew approximately 120 watts of power during sustained operation.
The model also handles 128K to 256K context windows without dropping speed, supporting extended workflows like code reviews across large codebases.
How does Edge0-145B compare to cloud models?
Edge0-145B scores competitively against frontier cloud models across six benchmarks covering coding, reasoning, knowledge, and instruction following. On DeepSWE 1.1, a software engineering benchmark, it scored 64.6 percent, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4.
On GPQA-Diamond, a graduate-level reasoning benchmark, it scored 90.4, close to DeepSeek-V4-Flash-0731's 90.8 and Claude-Opus-4.6 Max at 91.3. On IFBench, which tests instruction following, it scored 80.6, ahead of DeepSeek-V4-Flash-0731 at 79.2 and Claude-Opus-4.6 Max at 62.5.
On Terminal-Bench 2.1, an agent-focused benchmark that evaluates command-line task completion, the model scored 86.52, beating Qwen3.8-Flash-Next's 83.3, DeepSeek-V4-Flash-0731's 82.7, and Claude-Opus-4.6 Max's 78.2.
The average across all six benchmarks is 72.8, compared to 71.7 for Qwen3.8-Flash-Next. Edge0-145B achieves this at roughly 40 percent of the estimated memory footprint of Qwen3.8-Flash-Next, which Edge0 estimates at 70GB in standard 4-bit deployment.
The largest gains over Qwen3.8-27B, a mainstream on-device model baseline, are in agent and coding benchmarks: 53.1 percent improvement on DeepSWE and 18.5 percent on Terminal-Bench.
Scores are percentages. A dash means no public score is available.
Cloud models have no reported DeepSWE 1.1 score.
| Benchmark | Edge0-145B | Qwen3.8-Flash-Next | DeepSeek-V4-Flash-0731 | Claude-Opus-4.6 Max |
|---|---|---|---|---|
| DeepSWE 1.1 | 64.6 | 58.7 | 54.4 | \u2014 |
| Terminal-Bench 2.1 | 86.52 | 83.3 | 82.7 | 78.2 |
| GPQA-Diamond | 90.4 | 91.7 | 90.8 | 91.3 |
| IFBench | 80.6 | 81.3 | 79.2 | 62.5 |
How does it fit 145 billion parameters on a laptop?
Edge0 designed the model and its inference engine together to reduce the memory and data movement that make large models difficult to run locally. The design spans architecture, training, loading, and scheduling.
At the architecture level, the model shares large matrices across sparse experts. Each group of 64 experts shares two large matrices, while each expert retains a small set of parameters that captures its differences.
Previously, every expert stored a complete set of matrices. Of the model's 145 billion total parameters, only about 10 billion participate in computation on each inference step.
The inference pipeline runs across three storage tiers, overlapping read and computation to avoid serial waits. Frequently used hot experts remain in GPU memory.
An expert library of about 40GB sits in system memory, with experts brought into GPU memory as needed. About 48GB of vocabulary data resides on the SSD and is read on demand.
The CPU and GPU work in parallel: the GPU computes cached experts while the CPU handles the remaining work. Speculative decoding verifies multiple candidate tokens at once, spreading the cost of reading weights across more output tokens.
This approach allowed parameter counts to grow from tens of billions to more than a hundred billion without proportional increases in memory use, power consumption, or latency.
What can Edge0-145B do on a laptop?
The company demonstrated the model completing multi-step workflows entirely on-device. It produced a video explainer about the Tesla Roadster, handling technical research, scriptwriting, asset generation and editing within an hour.
It built a 3D Manhattan city scene in Blender, working from text and reference images, adjusting materials and lighting, and iterating on the rendered output.
Edge0 also showed the model creating playable games and building an original medieval capital city for a fantasy RPG, rendering a 20-second fly-through at game-trailer quality.
The model planned the city independently: city walls, more than a thousand old-town homes, a plaza fountain, a Gothic twin-tower facade, and stained-glass windows.
In a hardware demonstration, the model wrote a drone navigation program from one instruction: navigate through obstacles and find a red backpack.
It generated flight-control and visual-recognition programs locally, and the drone flew indoors, navigated through a dense row of obstacle poles, and located the red backpack beside a wall.
Each demonstration involved multiple rounds of planning, coding, execution, debugging, and visual inspection. Local inference meant those conversations, code revisions, and analyses incurred no cloud-model token charges.
What does this mean for local AI?
Previous on-device models stayed below 10 billion parameters. Apple Intelligence's on-device foundation model had approximately 3 billion parameters, while Google's Gemini Nano came in 1.8 and 3.25 billion sizes.
Even a 7 billion parameter model needed about 4GB of memory, and larger models faced a memory wall.
Apple acknowledged that its on-device model was not designed for agentic applications requiring general world knowledge. The prevailing assumption has been that on-device intelligence cannot match the cloud.
Edge0-145B changes that threshold. By bringing 145-billion-parameter capability to a laptop with 29GB memory, it removes the tradeoff between local privacy and cloud-scale capability.
Developers and power users can run coding agents, 3D workflows, and complex reasoning tasks without sending data to an API or paying per-token fees.
The model also compares favorably on efficiency: Edge0-145B's 72.8 mean score across six benchmarks comes with 29GB peak memory, while Qwen3.8-27B scores 66.1 with an estimated 18GB footprint and Qwen3.8-Flash-Next scores 71.7 with an estimated 70GB footprint.
The model weights, inference framework, and deployment tools are available on Hugging Face.
Common questions
How much memory does Edge0-145B need?
The model uses 29GB peak memory in the reported laptop setup, fitting within a 48GB laptop.
How fast is Edge0-145B?
It processes prompts at 1,100+ tokens per second and generates at 80+ tokens per second at approximately 120W power draw.
Is Edge0-145B truly open source?
Yes. Edge0 released the model weights, inference framework, and deployment tools as open source.
Sources & further reading
How this story was made
Written by Simon with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. The cover is an AI-generated editorial illustration, not a photograph or evidence of an actual event.
Our editorial standards





