FR
live
AI

Edge0-145B runs a 145-billion-parameter model on a single laptop

On 12 November 2026, AutoArk’s Edge0 team open-sourced Edge0-145B, a 145-billion-parameter mixture-of-experts model that runs locally on a MacBook Pro with 29 GB of peak memory and beats DeepSeek-V4-Flash on several coding benchmarks. Before adopting it, read the license terms and treat the scores as self-reported until they are reproduced.

A single open laptop on a dark desk emitting a thin amber light bar, a blurred wall of dim server racks receding in the background.

September 2026. Edge0 lands with a 35-billion-parameter model that runs on an iPhone 18 Pro Max. 12 November 2026. The team, part of AutoArk, open-sources the weights of Edge0-145B and Edge0-120B alongside its inference engine. Within two days, the GitHub repository passes 11,300 stars. Why it matters: for the first time, a 145-billion-parameter model claims to rival frontier cloud models without leaving the machine — a shift that changes the calculus for data residency and agent cost.

The memory wall that kept AI out of your machine

Before September 2026, on-device models stayed small for one reason: memory. Apple Intelligence’s on-device model ran around 3 billion parameters, Google’s Gemini Nano between 1.8 and 3.25 billion. A 7B model can already consume 4 GB of memory, and an aggressively compressed 35B approaches 20 GB — too much for a phone, barely enough for a laptop.

Edge0 breaks the wall through architecture rather than compression. The model is a latent MoE: of its 145 billion parameters, only ~10 billion participate in any given inference step. Each group of 64 experts shares two large matrices, and each expert keeps only a small set of parameters encoding its differences. The result: far fewer weights to store and move, without cutting the number of experts.

Loading is just as unusual. “Hot experts” stay in GPU memory, a roughly 40 GB library sits in system memory, and ~48 GB of vocabulary is read on demand from the SSD. The CPU and GPU work in parallel, with speculative decoding amortizing weight reads across multiple tokens. That is what produces, per the team, 29 GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode, and roughly 120 W of power on a MacBook Pro.

Cloud-level scores, to be read with care

The published numbers place Edge0-145B near frontier models on code and agent work:

  • DeepSWE 1.1: 64.6, ahead of Qwen3.8-Flash-Next (58.7) and DeepSeek-V4-Flash-0731 (54.4);
  • Terminal-Bench 2.1: 86.52, ahead of DeepSeek-V4-Flash-0731 (82.7) and Claude-Opus-4.6 Max (78.2);
  • GPQA-Diamond: 90.4, close to Claude-Opus-4.6 Max (91.3);
  • IFBench: 80.6, well ahead of Claude-Opus-4.6 Max (62.5).

The mean across six benchmarks reaches 72.8, versus 71.7 for Qwen3.8-Flash-Next — with a measured 29 GB footprint, about 40% of the ~70 GB estimated for a 4-bit deployment of the Qwen model. The largest gains over Qwen3.8-27B are in agentic work: +18.5% on Terminal-Bench and +53.1% on DeepSWE.

Two caveats are unavoidable. First, these scores are self-reported by Edge0, under its own evaluation conditions — the technical report and reproduction protocol are announced but not yet public. Second, the memory comparison mixes internal measurements (Edge0) and estimates (the competitor models, marked with an asterisk), so no definitive conclusion should be drawn before independent reproduction.

The limits Edge0 itself acknowledges

The honesty of the scoreboard deserves emphasis, because it draws the model’s real boundary. Edge0-145B tops out at 35.4 on HLE (Humanity’s Last Exam), versus 40.0 for Claude-Opus-4.6 Max — more than a 4-point gap on the most demanding general-knowledge evaluation. On AA-LCR, it reaches 79.3, roughly level with Claude (78.0). The message is clear: Edge0-145B dominates on code and agents, where its benchmarks surge, but it is not the model of encyclopedic knowledge.

The comparison with Qwen3.8-27B is just as instructive. The relative gains reach +18.5% on Terminal-Bench and +53.1% on DeepSWE, but only +1.3% on GPQA and +1.4% on IFBench. In other words, the model’s value is concentrated in what sets it apart — end-to-end agent work — and not in the pure reasoning tasks where smaller models already suffice.

That is also what mandates caution on the numbers: the evaluations are internal, and the technical report is not yet public. Deciding a local deployment on self-reported scores would be the same mistake as with a cloud vendor — the only valid proof remains independent reproduction on your own tasks.

What it changes for a CISO or an SRE

The real use case is not replacing Claude or GPT-6 everywhere: it is moving coding agents onto the machine. The published demos run from an explainer video produced in under an hour to a 3D Manhattan scene in Blender, a drone program written from a single instruction, and a PCB routed in KiCad then animated in Unreal Engine 5.

The economics are quantified: a Manhattan-scale city scene would typically cost $12,000–18,000 in labor, and a leading cloud model would exceed $150 in token charges for inference alone. Locally, those token fees disappear — hardware and electricity, of course, do not.

For a CISO, the benefit is elsewhere: code and data stay on the machine, with no account, no quota, and offline. That is the direct answer to data-residency and trade-secret constraints that cloud models impose — and the reason OpenAI, NVIDIA, and Qualcomm amplified the September 35B release.

Practically, the prerequisites are easy to check. The model needs a compatible laptop — typically a recent MacBook Pro or an RTX machine — with at least 29 GB of memory, and runs on macOS, Windows, and Linux. Weights come from Hugging Face, and inference goes through edge0-infer on the command line, a local API server, or the Edge0 Chat app. For a first run, a 48 GB machine leaves comfortable headroom before touching the SSD vocabulary tier.

There is one more engineering detail that makes the whole thing hold together: a router that predicts a step ahead. In a sparse model, the cost is not computing every expert but loading the right ones from slow storage in time. Edge0’s router anticipates which experts the next step will need and prefetches them, so the GPU never stalls waiting on the SSD. That is why the 80+ tokens/s figure is credible despite streaming ~48 GB of vocabulary from disk: the model hides storage latency behind prediction, the same trick that makes page-cache prefetching work in a conventional OS.

The verdict

If you have a machine with at least 48 GB and agentic code workloads that cannot leave the workstation, Edge0-145B deserves a serious test: the claimed footprint is realistic for a recent MacBook Pro, and the absence of token fees changes the math on long iterative tasks. If you are evaluating for production, sign nothing on the benchmarks: wait for independent reproduction and the technical report, check the license of the weights on Hugging Face, and measure real latency on your own tasks before generalizing. If your workloads stay in the cloud for compliance or simplicity, this is not the switch yet — but the trajectory is clear: a frontier model on your machine is no longer a lab promise.

References

The cyber brief, every Tuesday

The flaws that matter and the patches to apply, in a ten-minute read.

No spam. One-click unsubscribe.
read next

On the same topic

llama.cpp hits 1.0 and locks down the API the local AI ecosystem runs on

On 13 November 2026, llama.cpp shipped version 1.0.0 and applied semantic versioning to libllama and libggml, the libraries that carry local inference end to end. If you ship local AI in a product, pin a 1.x release through an official channel — Docker, Debian, or Homebrew — instead of vendoring your own build.

LiteLLM’s regex sandbox falls to bytecode rewriting and hands over a remote shell

On November 18, 2026, CISA added CVE-2026-40217 to its KEV catalog: LiteLLM’s guardrail-testing endpoint executes arbitrary code by rewriting Python bytecode, and the Docker image runs as root. Every LiteLLM gateway older than v1.83.10-stable must be upgraded, and the master key treated as a root secret.

← Back to the feed

Type at least two characters.

↑ ↓ navigate ↵ open esc dismiss