Edge0 open-sources Edge0-145B, a 145B model for laptops

Edge0 open-sources Edge0-145B, a 145B model for laptops

Edge0 announced that it is open-sourcing Edge0-145B and Edge0-120B, together with a unified inference framework and deployment tools. The company calls Edge0-145B the world's first open-source on-device AI model with 145 billion parameters, and says it brings coding, reasoning and agent capabilities that rival frontier cloud models to a personal computer. The post links the model weights, the inference framework and an Edge0 App. It says the full stack is released so developers can reproduce the results, evaluate or fine-tune the models and build them into products. A list of four ways the release addresses the on-device tradeoff is announced but not included in the text available. Specifications for Edge0-120B are not given either.

The post frames the problem as a memory wall. Before September 2026, it says, memory limits kept most on-device models below 10B parameters. Apple Intelligence's on-device model had about 3B parameters, and Google's Gemini Nano came in 1.8B and 3.25B versions. A 7B model could use about 4GB of memory, and even an aggressively compressed 35B model could need about 20GB. The post adds that Apple acknowledged its on-device model was not designed for agentic applications requiring general world knowledge. A scale chart in the post covers 67 open models of at least 1B parameters from 14 labs; Apple is excluded because its weights are not released.

Edge0 describes four layers of design. Architecture: each group of 64 experts shares two large matrices, while each expert keeps a small set of parameters that captures its differences; previously every expert stored a complete set. Of the 145B total parameters, only about 10B take part in computation on each inference step. Training: the model is trained in its on-device architecture from the start, rather than compressed or pruned after training a cloud model. Loading: reading and computation overlap across GPU memory, system memory and SSD. Frequently used "hot experts" stay in GPU memory, an expert library of about 40GB sits in system memory, and about 48GB of vocabulary data lives on the SSD and is read on demand. Scheduling: the GPU computes cached experts while the CPU handles the rest in parallel, and speculative decoding verifies several candidate tokens at once. These choices, the post says, support the reported decode rate of 80+ tokens/s (source: internal testing), with 28GB peak memory in the laptop comparison. For context, it cites its earlier Edge0-35B on an iPhone 18 Pro Max at 37+ tokens/s using about 2.5GB, and says these measurements apply to different models and hardware.

On benchmarks, Edge0 reports that Edge0-145B scores 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4. On GPQA-Diamond it scores 90.4, slightly below DeepSeek-V4-Flash-0731 at 90.8 and Claude-Opus-4.6 (Max) at 91.3. On IFBench it scores 80.6, versus 79.2 for DeepSeek-V4-Flash-0731 and 62.5 for Claude-Opus-4.6 (Max). On Terminal-Bench 2.1 it scores 86.52, against 83.3 for Qwen3.8-Flash-Next, 82.7 for DeepSeek-V4-Flash-0731 and 78.2 for Claude. On AA-LCR it scores 79.3 to Claude's 78.0. Against Qwen3.8-Flash-Next on a six-benchmark mean (DeepSWE 1.1, Terminal-Bench 2.1, HLE, GPQA-Diamond, IFBench, AA-LCR), Edge0-145B averages 72.8 to 71.7, with 29GB measured peak memory versus an estimated 70GB for standard 4-bit deployment of the baseline, about 40% of the estimated memory. The post says this enables local use on a 48GB laptop. Against Qwen3.8-27B, which it calls a mainstream on-device baseline, the gains are 18.5% on Terminal-Bench and 53.1% on DeepSWE, and the relative gains on HLE, GPQA and IFBench are 14.9%, 1.3% and 1.4%. The post states that the comparisons apply only to the named models and evaluations.

The post then shows demonstrations of local agent work, where the model plans, writes code, calls tools, debugs and checks screenshots. It says Edge0-145B produced an explainer video on how Tesla's Roadster could fly, covering research, scriptwriting, asset generation and editing, in under an hour with no cloud calls; the post says doing it manually would take a director, scriptwriter, 3D modeler, editor and physics expert a month. It says the model built a 3D scene of Manhattan in Blender within an hour, a job the post puts at roughly $12,000-18,000 and at least a month done by hand, with a leading cloud model still incurring more than $150 in token charges. It also describes a medieval capital for a fantasy RPG with city walls, more than a thousand old-town homes and a Gothic twin-tower facade, rendered as a 20-second fly-through; the post says an unnamed leading cloud model needed hundreds of tool calls and hundreds of millions of tokens for the same task. The last demonstration is hardware control: the model was given the instruction "Write a drone program that navigates through obstacles and finds a red backpack." and generated flight-control and visual-recognition programs for a real drone. The outcome of the drone test is not described in the available text.

Key facts

  • Edge0 open-sources Edge0-145B and Edge0-120B with an inference framework and deployment tools; Edge0-145B has 145B total parameters, of which about 10B are active per inference step.
  • Reported laptop figures: 80+ tokens/s decode and 28GB peak memory (29GB measured peak memory elsewhere in the post), enabling local use on a 48GB laptop.
  • Edge0 reports 72.8 mean across six benchmarks versus 71.7 for Qwen3.8-Flash-Next, whose roughly 70GB footprint is an estimate, not a measurement.
  • Individual scores: 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 80.6 on IFBench, and 90.4 on GPQA-Diamond, just below DeepSeek-V4-Flash-0731 (90.8) and Claude-Opus-4.6 (Max) (91.3).
  • All Edge0, Qwen, DeepSeek and Claude results come from Edge0's internal evaluations; the full evaluation setup and reproduction instructions are promised with a technical report.

Why it matters

Local models have been held back by memory. The post says most stayed below 10B parameters before September 2026, and that even a compressed 35B model could need about 20GB. Edge0 claims a 145B model that fits on a 48GB laptop by sharing matrices across experts, activating only about 10B parameters per step, and spreading weights over GPU memory, system memory and SSD. If the numbers hold, code and data can stay on the machine, and the post notes that local use avoids cloud token and API charges.

Who it affects

Developers who want to evaluate, fine-tune or embed a large model in their own products, since Edge0 releases weights, an inference framework and deployment tools. Also anyone who needs agent-style coding or reasoning to work offline or without sending code and data to a provider. The hardware bar is high: the post ties local use to a 48GB laptop, and charts a 64GB workstation as well.

How to use it

The post points to three things: the model weights, the inference framework and the Edge0 App. The reported setup keeps hot experts in GPU memory, an expert library of about 40GB in system memory and about 48GB of vocabulary data on the SSD. The exact laptop used for the speed and memory figures is not stated. No licence, download size or price is named in the available text.

How solid is it

Treat the results as the vendor's own. Edge0, Qwen, DeepSeek and Claude benchmark numbers all come from Edge0's internal evaluations, the decode rate is labelled internal testing, and the evaluation setup, hardware environment and reproduction instructions are still to come with a technical report. Baseline memory figures marked with an asterisk are estimates for standard 4-bit deployment, not measurements. The post gives 28GB peak memory in one place and 29GB in another without reconciling them. Its comparisons are limited to the named models and evaluations, and Claude-Opus-4.6 (Max) scores higher on GPQA-Diamond (91.3 versus 90.4).

Risks and caveats

The mean-score comparison against Qwen3.8-Flash-Next rests on an estimated 70GB baseline, so the "about 40% of memory" claim is against an unmeasured figure. The chart comparing quality and memory has a vertical axis that does not start at zero, as the post itself notes. Cost comparisons for the Manhattan scene and the medieval city rely on unnamed cloud models and the company's own estimates. The demonstrations are described by Edge0, and the drone outcome is not given in the available text. The laptop and phone speed figures belong to different models and hardware, so they are not directly comparable.

“Write a drone program that navigates through obstacles and finds a red backpack.”

— The instruction Edge0 says it gave Edge0-145B for its drone demonstration