Skip to content

Signals · Capability Shift

Edge0 puts a 145B-parameter model on a laptop, and agent work stops paying per token

Edge0 says its open 145B model runs locally in 29GB of memory, which cuts per-token costs for agent work and leaves the value with whoever writes the brief.

by ·4 min read·

New here? Start with the free AI Survival Kit →

Edge0 puts a 145B-parameter model on a laptop, and agent work stops paying per token

0:00 / 6:07

This voice is generated by AI.

Part of the guide Will AI take your job? How to find out for your role, in an afternoon

Edge0 puts a 145B-parameter model on a laptop, and agent work stops paying per token

Edge0 open-sourced Edge0-145B this week. The company says this 145-billion-parameter model runs on a compatible laptop, such as a MacBook Pro, according to its launch post on edge0.ai. Edge0 reports 29GB of peak memory and generation above 80 tokens per second. It also reports coding and agent scores that it says compete with specific frontier cloud models. Every benchmark in the post comes from Edge0's internal evaluations or from third parties it names. Edge0 says it is releasing reproduction instructions with its technical report. Until someone outside Edge0 reproduces the results, treat the numbers as a vendor claim. What matters for your work is the direction they point.

What moved

Edge0 writes that memory limits kept most on-device models below 10B parameters before September 2026. Apple Intelligence's on-device model had approximately 3B. Edge0-145B uses approximately 10B of its 145B parameters on each inference step. It shares large matrices across groups of 64 experts and loads the rest from system memory and SSD as needed.

Here are the scores Edge0 reports. On DeepSWE 1.1 it scores 64.6, ahead of Qwen3.8-Flash-Next at 58.7. On IFBench it scores 80.6, against 62.5 for Claude-Opus-4.6 (Max). On HLE, a hard reasoning test, Claude still leads 40.0 to 35.4. Keep that gap in view.

The cost line is what changes behaviour. Edge0's post says local inference carries no cloud-model token or API fees, data can stay on the computer, and core capabilities keep working offline. The company also notes that hardware and electricity still cost money. Its agent workflows take dozens of rounds of planning, running code, debugging and checking screenshots. Run locally, those rounds no longer add to a cloud bill.

The task layer gets cheaper again

Edge0's showcase is a video explainer about how Tesla's Roadster could fly. The company says the model handled the research, scriptwriting, asset generation and editing locally, and finished within an hour. It claims comparable manual work would have taken a director and scriptwriter, a 3D modeller, an editor and a physics expert a month.

Read that claim with care. It is a vendor demonstration, and Edge0 published no comparison of quality. It still shows where the line falls. The model did the execution. A person chose the subject and set the brief: make the principles understandable to a general audience. Edge0 does not say who checked the result before it was published. That layer keeps its value. AI does the routine work. You do the thinking.

Two shifts follow. First, the confidentiality objection to agentic AI gets weaker. Teams that kept client files, contracts and code away from cloud models for data-handling reasons now have a model that runs offline on a 48GB laptop, if their hardware qualifies. Second, extra attempts get cheap. When another run costs nothing per token, the scarce resource is the person who can judge which run was right.

Your Next Move

  • Run one bounded test. Pick a task you have kept off cloud models for confidentiality reasons, such as a code refactor, a contract summary or a data clean. If you have the hardware, run a local model on it. Record the time it saved and the errors you had to catch. Bring that record to your next planning conversation.
  • Re-sort your week. The Roadster demo folds four job descriptions into one workflow. List your own tasks and mark the ones that are execution an agent could run from a brief. The afternoon method for sorting your tasks and scoring your role gives you that list.
  • Own the brief and the sign-off. Before your next project starts, write down the standard the output must meet. The person who defines the finished state and approves it is the one the model multiplies.

I build systems like the one publishing this site. → Work with me

About the author

Jo

Jo runs The War Room: one signal a day on how AI is changing work, and what to do about it.

Sources

Get the Briefing

Want more intelligence like this?

The weekly briefing, free — and the AI Survival Kit with it.

Takes 30 seconds. No spam.

More in Signals

See all