Edge0 releases 145B open model that runs on a laptop
Edge0 launched two open-source models, 145B and 120B, built to run entirely on personal hardware.
· Last verified: 14 Nov 2026 (Edge0)

Summary
- Edge0 launched two open-source models, 145B and 120B, built to run entirely on personal hardware.
- The company says Edge0-145B peaks at 29GB of memory and generates over 80 tokens per second.
- If the internal scores hold, frontier-class work moves off the cloud, with no account or usage quota.
The latest
A 145-billion-parameter model that fits on a consumer laptop is now open-source. Edge0 announced Edge0-145B and Edge0-120B on November 12, 2026, alongside a unified inference framework and deployment tooling. The company describes the larger model as the first open-source on-device model at that parameter count, trained in its device-oriented architecture from the start rather than compressed after cloud training.
Details
- The architecture: Edge0 says the model holds 145 billion parameters total, with roughly 10 billion active per inference step. Its sparse-expert design groups experts in sets of 64 that share two large matrices, with each expert keeping only a small set of its own parameters — where previously every expert stored a full matrix set.
- The memory trick: Loading is split across three tiers, according to the company: hot experts stay in GPU memory, an expert library of roughly 40GB sits in system memory, and about 48GB of vocabulary data is read from SSD on demand. The GPU computes cached experts while the CPU handles the rest in parallel.
- The claimed numbers: Edge0 reports peak memory of 29GB, though its own comparison table lists 28GB. It also claims prefill above 1,100 tokens per second, generation above 80 tokens per second, typical power draw near 120 watts, and 128,000 to 256,000 tokens of context at comparable speed on long inputs.
- The benchmarks: On Edge0's internal evaluations, the 145B scores 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 35.4 on HLE, 90.4 on GPQA-Diamond, 80.6 on IFBench and 79.3 on AA-LCR, averaging 72.8 across six tests. The 120B averages 68.9 at the same 29GB.
- The comparison: Edge0 places its 145B above Qwen3.8-Flash-Next at 71.7 average and DeepSeek-V4-Flash-0731 at 70.1, both estimated to need roughly 70GB and 87GB. Claude-Opus-4.6 is listed as cloud-only and excluded from the average, scoring higher on HLE at 40.0 and GPQA at 91.3.
- The caveat: The company states that memory figures marked as estimates assume a standard 4-bit deployment and were not measured. Full evaluation setup and hardware details are to be published in a separate technical report. The blog post does not name a license for the released weights.
- The hardware: Edge0 says the model runs on 48GB laptops and 64GB workstations, citing MacBook Pro as an example, with day-one support for macOS, Windows and Linux across Apple M-series, NVIDIA RTX Spark, RTX 4090 and RTX 5090, plus one-click deployment tools and hardware guides.
- Availability: Weights are posted on Hugging Face as Edge0-145B-NVFP4 and Edge0-120B-NVFP4, with the inference framework and Edge0 App v0.2.0 on GitHub. Core app features run offline with no account and no usage quota, and support runs through Discord and GitHub.
- The demos: Edge0 showcases a Manhattan 3D scene built in Blender in under an hour, which it estimates would cost 12,000 to 18,000 dollars and a month manually, or over 150 dollars in token fees on a leading cloud model. Other demos include drone navigation software with no cloud call.
- The track record: Edge0 says its phone model Edge0-35B shipped September 10, 2026, peaking under 2.5GB and 37 tokens per second on iPhone 18 Pro Max. It reports 200 models across four families in a year, over 5 million devices in production, and more than 110,000 client organizations as of October 2026.
Background
Edge0 reports more than 15 million open-source downloads in three months and portfolio coverage of 95% of major chip platforms — while noting that coverage does not mean every device can run the 145B model.
Between the lines
Every performance and benchmark figure here is Edge0's own, measured in a setup the company has not yet published. The one external anchor it offers cuts both ways: it beats Claude-Opus-4.6 on Terminal-Bench and IFBench while trailing on HLE and GPQA. The hybrid deployment it cites suggests the pitch is displacement of most cloud calls, not all of them.
What's next
The promised technical report with full evaluation setup and hardware details is the next checkpoint, along with independent reproduction of the 29GB peak memory and 80 tokens per second claims on the named 48GB laptops.
Source: Edge0