What the 145B counts
The 145B total includes 51.2B parameters in a PLE n-gram embedding table and 84.6B in the expert pool. The card separately reports 10.12B active parameters per token: 10 of 512 experts per layer, plus one shared expert, across 48 layers. These describe different architectural quantities, not equivalent measures of computation or capability. The embedding table is a lookup structure, stored in fp8, memory-mapped from disk, and read rather than computed through. More than a third of the headline number is that table.
None of this makes the release fake. Sparse models have always separated total parameters from active ones, and a large embedding table can genuinely help a small active core. It does change what the number promises. A reader comparing 145B against a dense 70B model is comparing a mostly dormant store plus a 10B working set against 70B of working machinery. Edge0 calls the model the first open-source on-device model at that scale; the scale depends on which of the card's three numbers you count.
Where the memory actually lives
The second headline is 29GB of peak memory on one consumer GPU. The launch's 29GB peak claim and the card's residency figures are not reconciled in the materials reviewed. The card reports a 40.81 GiB Q3_K build and runtime residency of 23.14 GiB in VRAM plus 31.28 GiB in system RAM on the test machine, an RTX 5090 D with 24GB — 54.42 GiB across those two tiers. The launch post's design section describes the arrangement in prose: hot experts stay in GPU memory, an expert library of about 40GB sits in system memory, and about 48GB of vocabulary data stays on SSD and is read on demand.
Edge0 reports the model operating with one GPU, as a three-tier system that also needs a large block of RAM and fast storage, and the launch post itself frames the target as a 48GB laptop. Speed follows the budget. At full residency the card records 121.2 tokens per second of agent decode; constrained to 16 GiB of VRAM and 20 GiB of RAM, the same build records 44.9. The reported speeds differ substantially under the two memory budgets; they describe different memory configurations, and only one of them is in the headline.
The platform story needs the same care. The launch post claims day-one support for macOS, Windows and Linux, including Apple M-series chips. The model card says its build instructions target NVIDIA GPUs with CUDA 13 on Linux or Windows, and notes that its Apple M2 Ultra measurements do not establish macOS release availability. At the desk's check on November 13, the GitHub releases page's current release (v0.2.0) offered Windows desktop downloads only. Mac support is a claim in the announcement, not a documented install path.
The scores are the maker's
On capability, Edge0 reports a 72.8 mean across six benchmarks, against 71.7 for Qwen3.8-Flash-Next, the model it builds on, a DeepSWE 1.1 score of 64.6 versus 58.7, and a Terminal-Bench 2.1 score of 86.52. The launch post is plain about provenance: those results, including the rivals', come from Edge0's internal evaluations. The memory figure it beats, an estimated 70GB for a standard 4-bit deployment of the base model, is marked as an estimate that remains to be measured. The post also says the full evaluation setup and reproduction instructions are being released with a technical report, which had not appeared at launch.
That is the normal shape of a launch, and it is also the whole of the evidence so far. The weights are public, so the check is available to anyone with the hardware: download the Q3_K build, run a public benchmark harness, and publish the numbers next to the memory trace. Until someone outside Edge0 does that, the honest status of every score in this piece is "reported."
Open weights, reserved business
The company calls the release open-source. The code fits that description better than the model does. The inference framework is Apache-2.0. The weights carry the Edge0 Community Model License 1.0, stacked on the Qwen Community License 1.0 because the model builds on Qwen3.8-Flash-Next. The card's summary permits personal use, research, modification, internal use and commercial use of generated outputs. It reserves three things for a separate written license: selling or commercially bundling the model, operating a commercial model service, and offering an independent commercial coding or office assistant built on it.
Sources
- [1] Edge0, "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer," launch post, November 12, 2026Read source
- [2] Edge0, Edge0-145B Q3_K model card, Hugging Face (architecture, speed table, residency, license)Read source
- [3] Edge0, inference framework repository, GitHub (Apache-2.0 framework code)Read source
- [4] Edge0, GitHub releases page (v0.2.0 desktop downloads, Windows only, checked November 13, 2026)Read source
- [5] Daily AI Brief, "Edge0 releases 145B open model built to run on one consumer GPU," November 13, 2026 (secondary; figures attributed to the model card)Read source