Skip to content
Frontier Leaf FL markFLFrontierLeaf

A field journal of new AI developments.

Model updates

Developers run frontier 145B coding AI on laptops, cutting cloud costs and keeping data private

A new open-source model enables frontier-level coding and reasoning on personal laptops, removing cloud API costs and data privacy concerns for developers.

Depth
145BEdge0-145B runs locally on compatible laptops with reported peak memory usage of 29GB and decoding speeds above 80 tokens per second · reported
80 tokensEdge0-145B runs locally on compatible laptops with reported peak memory usage of 29GB and decoding speeds above 80 tokens per second · reported
1 outlet · newest source 84h ago
Developers run frontier 145B coding AI on laptops, cutting cloud costs and keeping data private

Key points

  • Edge0-145B runs locally on compatible laptops with reported peak memory usage of 29GB and decoding speeds above 80 tokens per second.
  • The model supports coding, reasoning and agent tasks, planning and executing code while debugging errors and checking visual results.
  • Local inference eliminates cloud token fees and allows sensitive data to remain on the device, maintaining core capabilities offline.

Edge0 has open-sourced its 145-billion-parameter model alongside a unified inference framework, enabling frontier AI capabilities on personal computers. This release aims to remove the tradeoff between model capability and hardware constraints that has defined on-device AI.This allows users to run powerful AI tools on their own laptops without paying for cloud services or sending data online.The release provides open-source weights and a unified inference framework for 145B-parameter local inference with 29GB peak memory.

Evidence: reportedEdge0-145B: Frontier AI Capabilities on Your Own Computer

aifleet-hn-ai

The evidence

According to a blog post published by Edge0, the Edge0-145B model is the first open-source on-device AI of this scale. The company reports that it runs on compatible laptops, such as MacBook Pro models, without requiring server infrastructure.reported Performance metrics cited in the post include a peak memory footprint of 29GB and decoding speeds exceeding 80 tokens per second.

Evidence: reportedEdge0-145B: Frontier AI Capabilities on Your Own Computer

aifleet-hn-ai

The decision

The release addresses four specific constraints in local AI deployment identified by the makers. Large models previously demanded excessive memory and power, while smaller models struggled with complex work. Edge0 states that their model processes long prompts, including dozens of pages of code, before generation begins. It sustains operation without throttling in the reported setup.reported

Evidence: reportedEdge0-145B: Frontier AI Capabilities on Your Own Computer

aifleet-hn-ai

The takeaway

Local inference removes recurring cloud-model token fees and dependence on network connectivity. Data handling remains on the computer, mitigating sensitive-data restrictions associated with cloud providers. However, the company notes that hardware and electricity costs still apply. The full stack is released so developers can reproduce results, evaluate or fine-tune the models, and integrate them into products.

Building on this

Developers can build applications that plan, write and run code autonomously. The model supports debugging errors and checking results visually for projects such as 3D scenes, games, hardware programs and PCB designs. Users can access the model weights, explore the inference framework or try the Edge0 App. The system maintains essentially the same speed on long inputs, supporting extended work sessions.

Evidence: reportedEdge0-145B: Frontier AI Capabilities on Your Own Computer

aifleet-hn-ai

Why it matters

  • Enterprise IT buyers can deploy large-language models internally without incurring per-token API fees or exposing sensitive data to third-party cloud providers.
  • Developers gain access to a unified inference framework that allows them to fine-tune and integrate 145B-parameter models into proprietary software products.
  • The reported performance of 80 tokens per second on laptop hardware suggests that high-latency cloud round-trips are no longer necessary for interactive coding assistants.
  • Offline capability ensures that core AI features remain functional during network outages, increasing reliability for field operations and secure environments.

Topics: models, open-source, local inference, AI agents, developer tools

Updated

Then / now / next

Then

On-device AI forced a tradeoff between model capability and hardware limits, with smaller models struggling on complex tasks.

Now

Edge0-145B delivers frontier-level coding and reasoning on laptops with 29GB peak memory and over 80 tokens per second.

Next

Developers will need to evaluate the open-source weights and inference framework to determine integration feasibility for their specific workflows.

Story timeline

  1. Then On-device AI forced a tradeoff between model capability and hardware limits, with smaller models struggling on complex tasks.
  2. 12 Nov 2026 Edge0-145B: Frontier AI Capabilities on Your Own Computer
  3. Now Edge0-145B delivers frontier-level coding and reasoning on laptops with 29GB peak memory and over 80 tokens per second.
  4. Next Developers will need to evaluate the open-source weights and inference framework to determine integration feasibility for their specific workflows.

The strongest doubt

Counterpoint from the sources

The company notes that hardware and electricity costs still apply, meaning local deployment is not free despite eliminating API fees.

For readers

Developers

  • Access model weights and inference framework to fine-tune and integrate 145B-parameter models into products.
  • Build agents that plan, write and run code while debugging errors and checking visual results for complex projects.

Business

  • Eliminate recurring cloud token fees and reduce dependence on network connectivity for core AI capabilities.
  • Keep sensitive data on local computers, mitigating privacy risks associated with cloud-based model providers.

Researchers

  • Reproduce reported results using the open-source weights and unified inference framework.
  • Evaluate the performance of 145B-parameter models on personal hardware without server infrastructure.

Try it

Plan and write code for a simple 3D scene

Good result: The model plans, writes and runs code, then checks the visual result for the 3D scene project.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Terms in this story

Inference framework

Software that enables AI models to process inputs and generate outputs on specific hardware.

Tokens per second

A metric measuring how quickly a model generates text, with higher numbers indicating faster output.

Open-source

Software whose source code is made available for use, modification and distribution by anyone.

Questions readers ask

What hardware is required to run Edge0-145B locally?

The model runs on compatible personal laptops, such as MacBook Pro models, with a reported peak memory requirement of 29GB.

Does the model support coding and agent tasks?

Yes, the model plans, writes and runs code, debugs errors and checks visual results for projects like 3D scenes and PCB designs.

Are there ongoing costs for using this model?

Local inference incurs no cloud-model token or API fees, though hardware and electricity costs still apply.

Connected stories

This storyOctomind agent harness enforces TOML policies for unattended code tasksOctomind agent harness enforces Google Workspace enables native Markdown editing in Docs for developers and AI agentsGoogle Workspace enables native Git-backed queues replace terminal juggling for AI coding agentsGit-backed queues replace terminNiCE Cognigy update prioritises agent reliability and operational visibilityNiCE Cognigy update prioritises

Reading receipt

  • Edge0-145B runs locally on compatible laptops with reported peak memory usage of 29GB and decoding speeds above 80 tokens per second.
  • The model supports coding, reasoning and agent tasks, planning and executing code while debugging errors and checking visual results.
  • Local inference eliminates cloud token fees and allows sensitive data to remain on the device, maintaining core capabilities offline.
Time spent: 0:00 · Depth: Full

Sources

  1. Hacker News: AI posts over 50 points

Related stories

Background reading