Edge0 open-sourced Edge0-145B on November 12, 2026, alongside a smaller Edge0-120B, a unified inference framework, and deployment tools. The startup describes Edge0-145B as the first open-source on-device AI model at 145 billion parameters, and says it brings coding, reasoning, and agent capabilities that rival frontier cloud models to consumer hardware.
The release landed on the same day it climbed to 219 points on Hacker News, where the prospect of a 100B-plus parameter model running without a server drew substantial attention from developers who follow local inference. For more context on this story, see our ongoing more AI stories.
The Numbers Behind the Laptop Claim
According to the company's announcement, Edge0-145B runs on a compatible personal laptop such as a MacBook Pro with the following reported results:
- 29GB peak memory, low enough to fit on high-end consumer machines without a dedicated server
- 1,100+ tokens per second prefill, letting the model process prompts spanning dozens of pages of code before generation begins
- 80+ tokens per second decode, which the company notes is faster than a person reads
- Approximately 120W typical power, sustaining operation without throttling in the reported setup
- 128K to 256K context, with essentially unchanged generation speed on long inputs
These figures come from Edge0's own published benchmarks rather than independent testing, and the company has released the full stack — weights, inference framework, and deployment tools — so developers can reproduce the measurements on their own hardware.
Why On-Device Models Have Stayed Small
The announcement includes useful context on why large models have historically been absent from laptops. Before September 2026, memory limits kept most on-device models below 10 billion parameters. Apple Intelligence's on-device foundation model had approximately 3 billion parameters, while Google's Gemini Nano shipped in 1.8B and 3.25B versions.
The constraint is physical: model weights must be read from memory during every generation step, competing for bandwidth with the operating system and other applications. Even an aggressively compressed 35B model could require around 20GB of memory, which is far beyond what phones could allocate.
Edge0's earlier release showed where the company's approach was heading. The Edge0-35B phone model reported peak memory below 2.5GB, delivering 37+ tokens per second on an iPhone 18 Pro Max. Edge0-145B has roughly four times the parameters of that model while its laptop configuration reports roughly double the decode speed at 28GB of memory — the company is careful to note these measurements apply to different models and hardware.
An Assistant That Finishes Work
Beyond raw speed, Edge0 positions the model as an agent rather than a chatbot. The announcement describes a system that plans, writes and runs code, debugs errors, and checks results visually for projects such as 3D scenes, games, hardware programs, and PCB designs.
The company reports that coding, reasoning, and instruction-following results are competitive with specific frontier cloud models, though its published comparisons have not yet been independently verified. As with any vendor-reported benchmark, developers will want to see third-party evaluations before drawing conclusions about frontier-level parity.
The Privacy and Cost Argument
The pitch rests on three advantages over cloud workflows. Local inference incurs no per-token API fees, sensitive data can remain on the device, and core capabilities keep working offline. For enterprises handling regulated data, and for developers in bandwidth-constrained environments, those properties can outweigh the raw capability of a hosted frontier model.
Hardware and electricity are not free, of course. Edge0 acknowledges this directly, noting that its power figures reflect a sustained ~120W draw — comparable to a high-performance laptop running a demanding workload, which has real implications for battery life and thermals.
What It Means for the Local AI Ecosystem
The announcement's accompanying analysis tracks 67 open models of at least 1B parameters from 14 labs intended for phones, laptops, and AI PCs, and the trajectory is steep. If the reported numbers hold up under independent review, Edge0-145B would represent a step change in what consumer hardware can host: a model class that until now demanded data center GPUs becoming available on machines people already own.
Competitors are moving in the same direction. Meta has shipped its Muse Glimmer open agentic model built to run on consumer GPUs, and Apple's Foundation Models framework has renewed developer interest in on-device inference. The difference is scale: nothing in the current open-weights ecosystem comes close to 145B parameters targeting laptop-class memory budgets.
The release also lands amid an intensifying debate about centralization. From privacy-focused browsers adding agentic payment flows to local-first inference frameworks, the industry's open-weight camp has spent 2026 arguing that capable AI does not need to live in a hyperscaler's data center. Edge0-145B is the most aggressive datapoint yet in that argument.
Developers can download the model weights, explore the inference framework, or try the Edge0 App through the company's site. Whether the benchmarks survive community scrutiny is the open question — but with the full stack released for reproduction, that scrutiny should not take long.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
