Edge0 Opens a 145B Model Built for One GPU

On-device AI developer Edge0 on Nov. 12 released the weights of Edge0-145B and a smaller Edge0-120B, two mixture-of-experts language models it says can run coding and agent tasks on a single consumer graphics card or a well-equipped laptop. The company calls Edge0-145B the world's first open-source on-device model with 145 billion parameters. All performance figures so far come from Edge0's own testing.

Key takeaways

• Edge0-145B activates about 10 billion of its 145 billion parameters per token and was trained natively at 4-bit precision, according to its model card.
• Edge0 reports 29GB peak memory and more than 80 tokens per second on a laptop, and 98.3 tokens per second in agent decoding on an RTX 5090 D with 24GB of video memory.
• The inference framework is Apache-2.0, but the weights carry a community license that bars commercial model services and resale without a separate deal.

How Edge0-145B fits a 145B model on one GPU

The Edge0-145B model card describes 48 layers with 512 experts each. For every token, 10 routed experts and one shared expert run, for 10.12 billion active parameters. The model supports a 256K-token context and lists Qwen3.8-Flash-Next as its base model.

Edge0 calls the design a "latent MoE." In its launch post, the company says each group of 64 experts shares two large matrices, so less data has to be stored and moved. Frequently used experts stay in GPU memory, an expert library of about 40GB sits in system memory, and roughly 48GB of vocabulary data is read from the SSD on demand.

On the RTX 5090 D, the NVFP4 build used 23.25GiB of video memory plus 38.27GiB of host RAM with memory residency enabled. The weight files are 40.81GiB for Q3_K and 47.79GiB for NVFP4.

Edge0-145B benchmarks against cloud models

Edge0 says the 145B model scored 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4. On Terminal-Bench 2.1 it reported 86.52, against 78.2 for Claude-Opus-4.6 (Max). It trails on harder reasoning: 35.4 on Humanity's Last Exam versus Claude's 40.0, and 90.4 on GPQA-Diamond versus Qwen's 91.7.

The company says those Edge0, Qwen, DeepSeek and Claude scores come from its internal evaluations. Memory figures for rival models, such as about 70GB for a standard 4-bit Qwen3.8-Flash-Next deployment, are Edge0's estimates rather than measurements.

Context: from 35B on phones to 145B on PCs

The release follows Edge0's first open-source drop in September, when the Edge0 GitHub repository published 35B and 8B preview models that stream expert weights from SSD. Edge0 says most on-device models stayed below 10 billion parameters before then.

Large open models have also become easier to run locally through tools such as llama.cpp, and sparse designs like DeepSeek's Flash line, which AI Newsway covered earlier, have pushed memory costs down.

License limits and what's next

The model card says Edge0's own contributions fall under the Edge0 Community Model License 1.0, with Qwen materials under the Qwen Community License. Personal, research and internal use are allowed, but selling the model, running a commercial model service or offering an independent commercial coding or office assistant requires a separate written license.

Platform support is also narrower than the launch post suggests. The model card says build instructions target NVIDIA GPUs with CUDA 13 on Linux or Windows, and that its Apple M2 Ultra measurements "do not establish macOS release availability." Edge0 says a technical report with full evaluation settings is being released.

FAQ

Can Edge0-145B run on a phone?
No. Edge0 says the 145B model targets suitably equipped laptops and PCs, while its 35B and 8B models are aimed at phones.

Is Edge0-145B free for commercial use?
Only partly. Internal use and commercial use of outputs are allowed, but commercial model services and resale need a separate license from Edge0.

Have the benchmark results been verified?
Not yet. All speed and benchmark figures come from Edge0's own testing.

References

• Edge0-145B: Open-Source On-Device AI with 145B Parameters — Edge0 (Nov. 12, 2026)
• Edge0/Edge0-145B-NVFP4 — Hugging Face model card
• GitHub - Edge0-AI/Edge0 — GitHub