Edge0 releases 145B open model it says runs on a single laptop

Edge0 reports a peak memory of 29GB on a compatible laptop, such as a MacBook Pro[1]. The same comparison lists prefill of 1,100+ tokens a second, decode of 80+ tokens a second, typical power of about 120W, and context from 128K to 256K[1]. Edge0 says only about 10 billion of the 145 billion parameters are active on each step[1].

The NVFP4 card measures a different machine, an RTX 5090 D with 24GB: 98.3 tok/s agent decode and 3,165.7 tok/s on a 16K prefill[2]. With residency on, that card held NVFP4 as 23.25 GiB of VRAM plus 38.27 GiB of host RAM[2]. The listed builds are Q3_K at 40.81 GiB and NVFP4 at 47.79 GiB[2].

Edge0 says each group of 64 experts shares two large matrices, and each expert keeps a small set of its own parameters[1]. Edge0 says it trained the model in that on-device architecture from the start, and that the release is not a compression or pruning step applied afterwards[1].

Edge0 pipelines the weights across GPU memory, system RAM and the SSD. Frequently used experts stay on the GPU, an expert library of about 40GB sits in system RAM, and about 48GB of vocabulary data is read from the SSD on demand[1]. The CPU and GPU run at the same time, and speculative decoding checks several candidate tokens together, Edge0 says[1].

The NVFP4 card lists 48 layers and 512 experts per layer, with 10 routed experts plus 1 shared expert active per token, 10.12 billion parameters active, and a 256K context[2].

Edge0 says the Edge0, Qwen, DeepSeek and Claude figures come from its own internal evaluations, and that a technical report with the full setup is being released. It presents the scores as percentages. They are not independently verified[1].

On DeepSWE 1.1, Edge0 reports 64.6 for Edge0-145B, 58.7 for Qwen3.8-Flash-Next and 54.4 for DeepSeek-V4-Flash-0731[1]. On Terminal-Bench 2.1, it reports 86.52, against 83.3 for Qwen3.8-Flash-Next, 82.7 for DeepSeek-V4-Flash-0731 and 78.2 for Claude-Opus-4.6 Max[1]. On GPQA-Diamond it reports 90.4, against 91.3 for Claude-Opus-4.6 Max, and on IFBench it reports 80.6[1].

Edge0 claims a mean of 72.8 across six benchmarks (DeepSWE 1.1, Terminal-Bench 2.1, HLE, GPQA-Diamond, IFBench and AA-LCR), against 71.7 for Qwen3.8-Flash-Next[1]. It estimates about 70GB for a standard 4-bit copy of that Qwen model, and says the estimate remains to be measured. Its own 29GB figure is a measured internal peak[1].

Both the NVFP4 and Q3_K cards set the licence to other, under the name edge0-community-model-1.0[2][3]. Edge0-specific contributions fall under the Edge0 Community Model License 1.0, only where Edge0 owns them or is permitted to grant the licence. Qwen materials stay under the Qwen Community License 1.0[2][3].

The cards say the Edge0 licence allows personal use, research, modification, noncommercial sharing, internal use, and commercial use of generated outputs[2][3]. A separate written licence is required to sell or bundle the model, to run a commercial model service, or to offer an independent commercial coding or office assistant[2][3]. Framework code retains an Apache 2.0 licence, and Strata retains an MIT licence[2][3].

Both cards list Qwen/Qwen3.8-Flash-Next as the base model[2][3]. Edge0's benchmark table compares Edge0-145B with that same model[1].

Edge0 says day-one support covers macOS, Windows and Linux, including Apple M-series chips and NVIDIA RTX Spark, RTX 4090 and RTX 5090[1]. The Edge0 App's core capabilities work offline, without an account or a usage quota, Edge0 says[1]. The model cards say build instructions target NVIDIA GPUs with CUDA 13 on Linux or Windows, and that Apple M2 Ultra measurements do not establish macOS release availability[2][3].

Edge0-35B, a phone model, was released on 10 September 2026[1]. The company says that by October 2026 it had more than 110,000 customer organisations, and that its open models passed 15 million downloads within three months[1]. It presents those as portfolio figures, and says the 145B results apply to Edge0-145B[1].

Edge0 also showed company demos: a Manhattan scene in Blender, a medieval city fly-through, a drone program that found a red backpack, a KiCad PCB taken into an Unreal Engine 5 animation, and an explainer video[1].

Edge0's case is a large local agent, with about 10 billion parameters active on each step of a 145 billion parameter model[1]. It says local inference adds no corresponding cloud-model token or API fee, and that code, documents and business records can stay on the machine. Core capabilities keep working offline. Features that depend on online information or services still need a connection[1].

The performance claims are self-reported. Memory, token rates and the benchmark mean are Edge0's, and the comparisons with Qwen, DeepSeek and Claude are internal[1]. Training compute, funding and an independent benchmark result are not in these sources.