ModelBrief · Edge0-145B · Topic Page
Last updated Nov 20, 2026

Edge0-145B Launches: Nov 12, 145B on One Consumer GPU

ModelBrief · Published 2026-11-20 · Fact-check 2026-11-20 · Reference: Edge0 official blog, Hugging Face model cards, GitHub Edge0-AI/Edge0

On November 12, 2026, the Edge0 team released two models at once: the 145B Edge0-145B and the 120B Edge0-120B, together with the inference framework edge0-infer, the deployment tooling edge0-strata, and the Edge0 App desktop application. The official blog post is titled "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer", and its pitch is a 145B model with frontier-level agent and coding ability running on your own computer. This article lists what was confirmed on launch day, then takes apart the two numbers in the headline — 29GB and 80+ tok/s — each of which leaves half the story out.

What was confirmed on launch day

ItemConfirmed
Release dateNov 12, 2026 (Thursday), official blog "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer"
DeveloperEdge0 (the AutoArk team; the official blog's contact address points to autoark.ai; company location not publicly disclosed)
Released the same dayEdge0-120B (the 120B sibling); release notes for inference framework edge0-infer v0.2.0 went up a day earlier, on Nov 11
PositioningOn-device agent / coding model: 145B parameters with only 10.12B active per token, built for local vibe coding and agent work
Base modelAlibaba's Qwen3.8-Flash-Next (Hugging Face model-card metadata; the official blog does not name Qwen)
WeightsOpen weights on Hugging Face in two builds: Q3_K 40.81 GiB and NVFP4 47.79 GiB (Edge0/Edge0-145B-q3k, Edge0/Edge0-145B-NVFP4)
Context256K tokens; KV in int8, 32K resident
Speed (vendor-measured, RTX 5090 D 24GB)NVFP4: 98.3 tok/s agent decode, 3,165.7 tok/s 16K prefill; Q3_K: 121.2 tok/s agent decode, 3,554.9 tok/s 16K prefill; the blog's headline figures are 80+ tok/s decode, 1,100+ tok/s prefill, ~120W typical power
LicenceWeights: Edge0 Community Model License 1.0, with the Qwen Community License 1.0 applying as well; framework code under Apache 2.0
PriceNo hosted API, no per-token billing; cost is local hardware and electricity only

Lineage: from a 35B phone model to a 145B laptop model

Edge0-145B did not come out of nowhere. According to the official blog, this on-device line began with Edge0-35B, released September 10, 2026 — a phone-scale model. The 145B and 120B scale the same idea up: an architecture that streams expert weights from SSD, from 35B to the 100B+ class. That lineage matters: Edge0-145B was never meant to be another cloud flagship. Its design goal is to fit cloud-scale parameters into consumer hardware — the architecture, the memory pipeline and the 4-bit training all serve that goal. See our architecture article.

Three things shipped on launch day: weights, framework, desktop app

This was not a weights-only drop — three pieces arrived together. First, the weights on Hugging Face, in Q3_K and NVFP4 builds. Second, the inference framework edge0-infer and deployment tooling edge0-strata on GitHub (Edge0-AI/Edge0, Apache 2.0); the v0.2.0 release notes went up on Nov 11, the local server speaks the OpenAI- and Anthropic-compatible Messages APIs, and the 145B weights are not bundled with the installer — they are downloaded separately from Hugging Face. Third, the desktop app Edge0 App, whose core capabilities work offline with no account and no usage quota (per the official blog). Having all three on day one is the biggest difference between Edge0-145B and a research-only model: it could be installed and run on launch day.

The two headline numbers: 29GB and 80+ tok/s each leave half out

29GB is GPU-side peak memory, not whole-machine demand. The official blog headline reads "29GB peak memory", "runs on a compatible personal laptop, such as a MacBook Pro", and "enabling local use on a 48GB laptop". But the model card's own measured residency (RTX 5090 D 24GB) tells a different story: NVFP4 needs 23.25 GiB VRAM + 38.27 GiB host RAM, about 61.5 GiB in total; Q3_K needs 23.14 GiB VRAM + 31.28 GiB host RAM, about 54.4 GiB in total — plus an expert library of about 40GB in system memory and vocabulary data of about 48GB on SSD. "One consumer GPU" is true, but it is followed by a large helping of system RAM and SSD — the laptop headline runs ahead of the shipping docs. See our install guide.

80+ tok/s depends on the build. The blog's headline figures are 80+ tok/s decode and 1,100+ tok/s prefill; the model card measures per build: NVFP4 at 98.3 tok/s agent decode and Q3_K at 121.2 tok/s, with 16K prefill at 3,165.7 and 3,554.9 tok/s respectively (RTX 5090 D). Quote speed with the build name and the hardware — a single number misleads.

The vendor's wording, and our label: "open-source" needs quotation marks

The official blog calls this release "open-sourcing" and calls Edge0-145B "the world's first open-source on-device AI model with 145 billion parameters". That wording does not hold: the weights are actually under the Edge0 Community Model License 1.0 (dated Nov 12, 2026) — a custom community licence under which commercial services need a separate written licence — and the Qwen Community License 1.0 applies as well. Only the framework code is Apache 2.0. The accurate term is "open weights", not "open source". This site uses "open weights" throughout, and our licence article works through the licence clause by clause — see licence and cost.

The first week after launch: many explainers, zero independent verification

Between Nov 13 and 15, a wave of explainers and critical pieces arrived: groundtruth.day and oguzhan.co each flagged the 29GB and Mac-support discrepancies and took apart the licence, with freshsynapse, explainx.ai and tokenstrom following, and newtoken.link publishing a critical piece in Chinese; Signal Sentry published a scorecard. What needs saying is that all of these pieces re-analyse the vendor's numbers — none of them actually ran the model. As of Nov 20, 2026, there is still no independent benchmark score of any kind (no AA, LiveBench or LMArena listing), and the technical report is still unpublished. The first week's consensus is "plenty of material, zero verification" — see our benchmarks article.

ModelBrief's take: This is a genuine release: weights, framework and app all shipped together, downloadable and installable on Nov 12 — not a paper announcement. But both headline selling points need discounting: 29GB is GPU-side peak only, with about 61.5 GiB for the whole machine; and "open-source" is really open weights under a custom licence. Read those two points first, then decide whether it is worth installing.