What is Edge0-145B?
Edge0-145B is an open-source 145B-parameter model built to run locally on laptops and desktops, with reported 29GB peak memory, 80+ tokens/s decoding, and 128K–256K context.
Updated ·
Edge0-145B: Open-Source 145B On-Device AI Model
Announcement date:
Edge0-145B is described as the world's first open-source on-device AI model with 145 billion parameters. Its key claim is architectural: although the model holds 145B parameters in total, only about 10B participate in computation on each inference step, which is why it can reportedly run locally on a compatible laptop with roughly 29GB peak memory and decoding above 80 tokens/s. This profile covers what the model is, the tasks it suits, how to get started, documented limits, and the benchmark and deployment details published with the release.

What Edge0-145B is
Edge0-145B is an open-source large language model with 145 billion total parameters, positioned as an on-device model rather than a data-center-only system. The release frames it as the first open-source on-device model at this parameter scale.
The design goal is to keep frontier-style capability on a personal computer. Reported figures include 29GB peak memory on a compatible laptop, decoding above 80 tokens/s, prefill of 1,100+ tokens/s, and typical power consumption of approximately 120W. These are vendor-reported numbers from the announcement, not independently verified results.
- 145B total parameters, open-source release
- Positioned as an on-device model for local use
- Reported 29GB peak memory and 80+ tokens/s decoding
- Reported 1,100+ tokens/s prefill and roughly 120W typical power
Why a 145B model can run on a laptop
The central mechanism is sparse computation. Of the model's 145B total parameters, only approximately 10B participate in computation on each inference step, so the compute cost per token is far below what a dense 145B model would require.
Memory is handled through sharing and tiering. Each group of 64 experts shares two large matrices, while each expert keeps a small set of parameters that captures its differences. An expert library of approximately 40GB sits in system memory, with experts brought into GPU memory as needed, and approximately 48GB of vocabulary data resides on the SSD and is read on demand.
This is why the model can fit on consumer hardware, but it also explains the dependency on fast memory, SSD, and GPU transfer paths. The reported speeds assume a compatible machine; weaker storage or memory subsystems may behave differently.
- Only about 10B of 145B parameters compute per step
- 64-expert groups share two large matrices; experts keep small differentiating parameters
- ~40GB expert library in system memory, loaded into GPU memory as needed
- ~48GB vocabulary data on SSD, read on demand
Useful tasks and context handling
Because the model is designed for local execution, it fits workflows where data should stay on the machine or where cloud round-trips are undesirable. The published benchmarks emphasize coding, reasoning, instruction following, and terminal-style agent work, which suggests those as the natural first areas to evaluate.
The model supports a 128K–256K context window and is reported to maintain essentially the same speed on long inputs, which matters for extended work such as reading large codebases, long documents, or multi-step agent sessions. Treat the long-context speed claim as reported rather than guaranteed on every configuration.
- Local coding assistance and codebase reading
- Reasoning and question answering tasks
- Instruction-following and structured output work
- Terminal and agent-style multi-step tasks
- Long-document and long-context sessions up to the supported window
Published benchmark scores
The release reports 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4. It also reports 90.4 on GPQA-Diamond, 80.6 on IFBench, and 86.52 on Terminal-Bench 2.1.
On GPQA-Diamond the announcement places it close to DeepSeek-V4-Flash-0731 at 90.8 and Claude-Opus-4.6 (Max) at 91.3. These comparisons come from the same announcement, so they should be read as vendor-reported context rather than a neutral third-party evaluation.
- DeepSWE 1.1: 64.6
- GPQA-Diamond: 90.4
- IFBench: 80.6
- Terminal-Bench 2.1: 86.52
Getting started and platform support
Day-one support covers macOS, Windows, and Linux, including Apple M-series chips and NVIDIA platforms such as RTX Spark, RTX 4090, and RTX 5090. The release includes one-click deployment tools and hardware guides, with model weights on Hugging Face and the inference framework on GitHub.
A practical starting point is to check the hardware guide for your machine before downloading anything, since the memory tiering depends on system RAM, GPU memory, and SSD speed together. If your device is far below the supported class, plan on a smaller model or a cloud endpoint instead.
- Supported OSes: macOS, Windows, Linux
- Supported hardware includes Apple M-series and NVIDIA RTX Spark, RTX 4090, RTX 5090
- One-click deployment tools and hardware guides included
- Weights on Hugging Face; inference framework on GitHub
Documented limits and pricing status
The clearest limit is hardware. A very low-compute device cannot currently run a 145B model locally and needs a smaller model or cloud access. The reported 29GB peak memory and 80+ tokens/s figures apply to a compatible laptop, not to arbitrary machines.
Pricing for hosted access or commercial terms is not specified in the supplied release details, so it should be treated as unknown here. The model itself is described as open-source, with weights published on Hugging Face and the framework on GitHub.
- Very low-compute devices cannot run it locally today
- Performance figures assume compatible hardware
- Hosted or commercial pricing is not stated in the available details
- Open-source weights and framework are published
Questions
What hardware and memory does Edge0-145B need to run locally?
The release reports running on a compatible laptop with 29GB peak memory, and lists day-one support for macOS, Windows, and Linux, including Apple M-series chips and NVIDIA platforms such as RTX Spark, RTX 4090, and RTX 5090. An expert library of about 40GB sits in system memory and about 48GB of vocabulary data is read from the SSD on demand, so RAM, GPU memory, and storage speed all matter.
How fast is Edge0-145B, and what power does it draw?
Reported figures are decoding above 80 tokens/s, prefill of 1,100+ tokens/s, and typical power consumption of approximately 120W. The model is also said to maintain essentially the same speed on long inputs within its 128K–256K context window. These are vendor-reported numbers for compatible hardware.
Why does a 145B model only compute about 10B parameters per step?
The architecture is sparse: of the 145B total parameters, only approximately 10B participate in computation on each inference step. Each group of 64 experts shares two large matrices while each expert keeps a small set of differentiating parameters, and experts are brought into GPU memory as needed. That keeps per-token compute low while the full parameter set stays available.
Can a low-compute device run Edge0-145B locally?
No. A very low-compute device cannot currently run a 145B model locally and needs a smaller model or cloud access. The reported local performance assumes a compatible laptop or desktop in the supported hardware class.
Original sources
Tool Info
Has this tool changed? Help us keep it up to date
Edge0-145B Reviews
Loading reviews…