The on-device agent / coding model released November 12, 2026 by the Edge0 team (AutoArk): 145B total parameters with only 10.12B active per token, built on Alibaba's Qwen3.8-Flash-Next, with weights open on Hugging Face. The official headline says "29GB peak memory — runs on one consumer computer" — but the model card's own measurements tell a different story: about 61.5 GiB for the whole machine on NVFP4, and no Mac support at launch. On benchmarks it wins big on two coding scores and trails its base on three knowledge scores, with a mean lead of just 1.1 points — and all six scores are vendor-reported, with no independent verification yet. This page collects the official specs, the install requirements, the licence, and the first coverage in English and Chinese — with unverified claims clearly marked.
What is confirmed: this site's first topic on "a 145B in one consumer computer" — "open weights, runs locally" is the real selling point, while the "29GB, open-source, MacBook Pro" claims all need discounting.
Where it stands among base derivatives: two coding wins, a mean lead of just 1.1 points; the biggest difference is not the score but the deployment shape — it is the only one that runs locally.
| Model | Mean (6, vendor-reported) | DeepSWE 1.1 | Terminal-Bench 2.1 | Weights | Cost |
|---|---|---|---|---|---|
| Edge0-145B | 72.8 | 64.6 | 86.52 | Open weights (custom licence) | Local hardware + power, no API price |
| Qwen3.8-Flash-Next (base) | 71.7 | 58.7 | 83.3 | Qwen community licence | Cloud model |
| DeepSeek-V4-Flash-0731 | 70.1 | 54.4 | 82.7 | — | Cloud model |
All benchmark scores are vendor-reported (Edge0 internal evaluations) with no independent verification as of Nov 20, 2026. Edge0-145B's mean lead over its base is only +1.1, and three scores (HLE, GPQA-Diamond, IFBench) sit below the base; the mean lead over DeepSeek-V4-Flash is +2.7. DeepSeek-V4-Flash-0731 is a July 31 build.
Six benchmarks, grouped: Edge0-145B's lead concentrates on the two coding tasks, DeepSWE and Terminal-Bench, while HLE, GPQA-Diamond and IFBench trail the base.
Short on time? Start with these three: the benchmark tally, the licence analysis, and the install guide — the three places the headline misleads most.
Eight days after launch, the critical pieces arrived faster than any independent verification: three key critiques landed on Nov 13–15, flagging the 29GB, Mac and licence discrepancies, while independent scores remain at zero.
Edge0-35B launches (a phone-scale model; per the official blog, the starting point of the 145B line).
edge0-infer v0.2.0 release notes go up (GitHub): Windows x64/arm64 packages, the local server speaks the Anthropic Messages API; the notes say the 145B weights are not bundled with the installer and must be downloaded separately from Hugging Face.
Edge0-145B and Edge0-120B launch: the official blog "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer" goes up, weights land on Hugging Face (Q3_K and NVFP4 builds), framework and deployment tooling on GitHub, and the Edge0 App ships; the blog calls it "the world's first open-source on-device AI model with 145 billion parameters", with 29GB peak memory, runnable on a MacBook Pro.
groundtruth.day publishes a critical piece flagging the 29GB and Mac-support discrepancies and taking apart the licence; Signal Sentry publishes its scorecard, stating "No independent results yet".
freshsynapse and explainx.ai publish explainers; newtoken.link publishes the critical Chinese piece "Edge0's 145-billion-parameter model cannot run on Mac at launch"; the After Intelligence newsletter (edition 070) leads with Edge0-145B as a frontier item, alongside Nex-N2.6.
oguzhan.co publishes a deep critical piece, "what the 29GB headline leaves out": the NVFP4 whole-machine figure of about 61.5 GiB, the Mac gap, and the licence detail; a YouTube deep-dive appears (commentary on the vendor numbers, not an independent run); Hacker News discussion follows (the "open-weight rather than open-source" debate).
This site completes its research: the six benchmarks, the licence text and the install requirements are confirmed, and the verdict is that the material supports a hub + 6 articles — the topic is approved.
This site's topic page goes live (this page). As of today: the technical report is still unpublished, AA, LiveBench and LMArena have no Edge0-145B listing, and OpenRouter has not listed it.
As published by the vendor (Hugging Face model cards, read directly by this site on Nov 16, 2026; official blog checked the same day).
There is no cloud API, so trying it means local routes only: the command line, the desktop app and a direct weight download all work; the Mac route does not work at launch.
Get edge0-infer v0.2.0 (release notes Nov 11, 2026) from GitHub Edge0-AI/Edge0; the local server speaks the OpenAI- and Anthropic-compatible Messages APIs, and tool calling goes through it. Weights are not bundled with the installer — download them separately from Hugging Face. Requirements: NVIDIA GPU (compute capability 7.5+), 16 GiB+ VRAM, 20 GiB+ system RAM, CUDA Toolkit 13, Linux or Windows. See our install guide.
Per the official blog, the core capabilities work offline with no account and no usage quota. But the v0.2.0 Edge0 Chat desktop packages are Windows-only — "download the app and open it" currently works for Windows users only.
Q3_K (Edge0/Edge0-145B-q3k, 40.81 GiB) or NVFP4 (Edge0/Edge0-145B-NVFP4, 47.79 GiB). Default to Q3_K: smaller file, less host RAM, faster (121.2 vs 98.3 tok/s decode). The download is free; the licence is in section 08.
Input supports text + image: the model folder carries a vision encoder, mmproj-BF16.gguf, but it needs the strata-vision helper (not built from the Edge0 tree) — check its status during setup; output is text. Tool calling goes through the local OpenAI- / Anthropic-compatible server.
Also: the 120B sibling Edge0-120B launched the same day, also in Q3_K and NVFP4 builds; its full specs and benchmark table are not yet published by the vendor — see section 13.
There is no per-token price, so there is no "$ per 1M tokens" table — Edge0-145B's ledger has two lines only: hardware and electricity. Prices checked Nov 16, 2026 (model cards and official blog, read directly).
| Cost item | Edge0-145B | Cloud API models (comparison) |
|---|---|---|
| API calls (per 1M tokens) | None (no hosted API, no OpenRouter listing) | Billed per use |
| Weight download | Free (open weights on Hugging Face) | — |
| Framework and tools | Free (Apache 2.0) | — |
| Hardware (one-off) | 16GB+ VRAM GPU, 20 GiB+ system RAM, CUDA 13; measured whole-machine demand about 61.5 GiB (NVFP4) | None |
| Electricity (marginal) | About 120W typical power (vendor figure) | None |
The official pitch of "zero token cost" is true — but it tells half the story: the marginal cost is zero, the fixed cost is not. Before running Edge0-145B you need a machine that meets the bar — a 16GB+ VRAM GPU, 20 GiB+ system RAM, CUDA 13, and about 40–48 GiB of disk; measured whole-machine memory demand runs about double the headline 29GB (about 61.5 GiB for NVFP4). For someone who already owns that machine, the marginal cost is indeed only electricity (about 120W typical power); for someone who does not, the first bill is hardware, not tokens.
Compare it with cloud API models this way: cloud bills per use with zero fixed cost; Edge0-145B pays its fixed cost once and only electricity afterwards — the heavier your usage, the better the local route looks; light usage is cheaper in the cloud. Do not read "no token bill" as "free".
Official pages: Edge0 official blog · Hugging Face Q3_K model card · Hugging Face NVFP4 model card
All from Nov 13–16: nobody actually ran the model — every piece re-analyses the vendor's numbers. Every entry carries a verdict badge; "Read on this site" opens our rewritten deep-dive article, with the original source at the end.
Source: Signal Sentry, Nov 13–15, 2026
The original states it plainly: "No independent results yet. They appear here as Epoch AI and LMArena add this model." It also confirms OpenRouter does not list the model. The only "verification status" list from the first week: the benchmark, memory and speed figures all come from Edge0's own testing so far.
Source: oguzhan.co, Nov 15, 2026
Works out the NVFP4 whole-machine figure of about 61.5 GiB (23.25 GiB VRAM + 38.27 GiB host RAM), flags the Mac-support gap, takes apart the licence, and reproduces the six-benchmark table with the vendor label. This page's memory and licence wordings agree with it; the piece has no independent run either.
Read on this site → (original ModelBrief rewrite)
Source: groundtruth.day, Nov 13, 2026
Flagged the 29GB and Mac-support discrepancies the day after launch, with licence detail, and noted that the blog table's Claude-Opus-4.6 comparison column is outdated. One of the earliest and most carefully checked critical pieces.
Read on this site → (original ModelBrief rewrite)
Source: freshsynapse (Nov 14), explainx.ai
Both explainers actually checked the Hugging Face model cards and the GitHub repository, confirming the Q3_K / NVFP4 builds, the CUDA 13 requirement and the 16GB VRAM floor; explainx.ai states the "open weights, not open source" framing explicitly. Their value is in fact-checking, not new measurement.
Read on this site → (original ModelBrief rewrite)
Source: newtoken.link, Nov 14, 2026
Headlined "Edge0's 145-billion-parameter model cannot run on Mac at launch": it makes three points — the Mac controversy, the mean lead of only +1.1 over the base, and the licence discount. The only Chinese piece with substantive criticism; no tier-1 Chinese tech media (Jiqizhixin, Xinzhiyuan, QbitAI, IT Home) has reported on it.
Read on this site → (original ModelBrief rewrite)
Source: After Intelligence newsletter, edition 070, Nov 14–15, 2026
Leads with Edge0-145B as a frontier item, alongside Nex-N2.6. It shows the model had entered English newsletters' view within three days of launch; the newsletter does no independent measurement, so do not cite it as verification.
Source: YouTube, Nov 15, 2026 (about 7 minutes, Edge0 + Gemini 4 leaks)
Video commentary on the vendor numbers, with no real hardware run and no independent measurement. Cite it only as "video commentary", never as "independent testing".
The six articles below are original ModelBrief rewrites and reorganization — restructured with added conclusions and buying advice, with facts and numbers checked against the official sources. Read them on this site; each ends with its original source.
Edge0-145B has no score on any independent board yet: no AA, LiveBench or LMArena listing. The six scores below all come from the official blog's internal evaluations; the technical report is still unpublished and no reproduction instructions are out. Check the conditions and wording before the scores.
| Benchmark | Edge0-145B | Qwen3.8-Flash-Next (base) | DeepSeek-V4-Flash-0731 | Claude-Opus-4.6 (Max) |
|---|---|---|---|---|
| DeepSWE 1.1 | 64.6 | 58.7 | 54.4 | — |
| Terminal-Bench 2.1 | 86.52 | 83.3 | 82.7 | 78.2 |
| HLE | 35.4 | 35.9 | 33.8 | 40.0 |
| GPQA-Diamond | 90.4 | 91.7 | 90.8 | 91.3 |
| IFBench | 80.6 | 81.3 | 79.2 | 62.5 |
| AA-LCR | 79.3 | 79.1 | 79.7 | 78.0 |
| Mean (6) | 72.8 | 71.7 | 70.1 | — |
Full tally: our benchmarks article →
| Dimension | Edge0-145B | Qwen3.8-Flash-Next |
|---|---|---|
| Mean (6, vendor-reported) | 72.8 | 71.7 (+1.1) |
| DeepSWE 1.1 | 64.6 | 58.7 (+5.9) |
| Terminal-Bench 2.1 | 86.52 | 83.3 (+3.22) |
| HLE / GPQA-Diamond / IFBench | 35.4 / 90.4 / 80.6 | 35.9 / 91.7 / 81.3 (higher on all three) |
| Deployment | On-device: one consumer GPU + large memory | Cloud model |
| Weights | Open weights (Edge0 community licence) | Qwen community licence |
| Context | 256K | — |
What the rebuild bought is direction, not breadth: +9.1 points combined on the two coding tasks, −2.5 combined on the three knowledge and instruction-following tasks, a net mean gain of 1.1. See our comparison article.
| Dimension | Edge0-145B | DeepSeek-V4-Flash-0731 |
|---|---|---|
| Mean (6, vendor-reported) | 72.8 | 70.1 (+2.7) |
| DeepSWE 1.1 | 64.6 | 54.4 (+10.2) |
| Terminal-Bench 2.1 | 86.52 | 82.7 (+3.82) |
| GPQA-Diamond / AA-LCR | 90.4 / 79.3 | 90.8 / 79.7 (higher on both) |
| Deployment | On-device, runs locally | Cloud model |
The conclusion against DeepSeek matches the one against the base: the lead concentrates on the two coding scores, while GPQA-Diamond and AA-LCR still trail. DeepSeek-V4-Flash-0731 is a July 31 build. See our comparison article.
Edge0-120B launched the same day as the 145B — the smaller tier of the same line, also in Q3_K and NVFP4 builds. But as of Nov 20, 2026, neither the official blog nor the model cards publish the 120B's full specs or a six-benchmark comparison table — this site makes no quantified claim about "how much stronger the 145B is than the 120B". Some third-party pieces quote Terminal-Bench figures for the 120B, but the numbers contradict each other across sources (75.3 vs 81.3) — none of them is used until officially confirmed.
Most confusion about Edge0-145B comes from six places: the licence, the 29GB figure, the Mac, the "better across the board" claim, the API, and the Claude comparison — clarified one by one below.
The weights are under the Edge0 Community Model License 1.0 (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". The official blog's "open-sourcing" wording does not match the licence text.
29GB is GPU-side peak memory, not whole-machine demand. The model card's measured residency: NVFP4 needs 23.25 GiB VRAM + 38.27 GiB host RAM (about 61.5 GiB total), Q3_K needs 23.14 GiB VRAM + 31.28 GiB host RAM (about 54.4 GiB 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 official blog claims day-one support for macOS / Apple M-series and markets it with a MacBook Pro as the example; but the GitHub README's platform table lists no macOS support for the 145B/120B models, the build instructions target NVIDIA GPUs with CUDA 13 on Linux or Windows, and the Edge0 App is Windows-only. The 57.0 tok/s on M2 Ultra is a Metal speed measurement, not proof of availability.
The mean lead over the base is only 1.1 points (72.8 vs 71.7), and three scores sit below the base (HLE, GPQA-Diamond, IFBench). The lead concentrates on DeepSWE (+5.9) and Terminal-Bench (+3.22) — a directional lead on two coding tasks, not a general one.
Edge0-145B has no hosted API, no per-token price, and OpenRouter does not list it. Calls go only through the local edge0-infer server (OpenAI- / Anthropic-compatible Messages APIs), and the cost is hardware and electricity only.
The blog table sets Edge0-145B's Terminal-Bench 2.1 (86.52) against Claude-Opus-4.6 (78.2), but commenters have already pointed out that this comparison is outdated — the Opus 4.6 figure is not a same-period, same-protocol measurement. State the timeliness problem before citing it.
November 12, 2026 (Thursday); the 120B sibling Edge0-120B launched the same day, and the edge0-infer v0.2.0 release notes went up a day earlier, on Nov 11. See section 05 for the timeline.
Edge0 is the on-device AI brand of the AutoArk team; the official blog's contact address points to autoark.ai. The company's location is not publicly disclosed, and this site found no credible source for it, so no country is stated on this page.
No — it is open weights. The weights are under the Edge0 Community Model License 1.0 (Nov 12, 2026), with the Qwen Community License 1.0 applying as well; only the framework code is Apache 2.0. Personal use, research, modification, non-commercial sharing, internal business use and commercial use of generated outputs are free; selling the weights, operating a commercial model service for third parties, or offering an independent commercial coding or office product needs a separate written licence. See our licence article.
Alibaba's Qwen3.8-Flash-Next (Hugging Face model-card metadata; the official blog does not name Qwen). The routed experts in all 48 layers were rebuilt as latent MoE and post-trained on Edge0's own data — a derivative, not a new base model.
29GB is GPU-side peak memory only. The model card's measured residency: NVFP4 needs about 61.5 GiB for the whole machine (23.25 GiB VRAM + 38.27 GiB host RAM), Q3_K about 54.4 GiB, plus an expert library of about 40GB in system memory and vocabulary data of about 48GB on SSD. The bar is an NVIDIA GPU (compute capability 7.5+), 16 GiB+ VRAM, 20 GiB+ system RAM, CUDA 13. See our install guide.
Not at launch. The official blog claims day-one macOS support, but the GitHub README's platform table lists no macOS support for the 145B/120B models, and the Edge0 App is Windows-only. The 57.0 tok/s on M2 Ultra is a Metal speed measurement, not proof of availability. Wait until the README adds macOS support.
There is no per-token billing: no hosted API, no OpenRouter listing — the cost is local hardware and electricity only (about 120W typical power, per the vendor). Weight download and framework are free. See our licence and cost article.
Much better on coding and agents: DeepSWE 1.1 up 5.9 points, Terminal-Bench 2.1 up 3.22 points; but three scores sit below the base (HLE, GPQA-Diamond, IFBench), and the mean lead is only 1.1 points (72.8 vs 71.7). All six scores are vendor-reported, with no independent verification yet. See our benchmarks article.
On Hugging Face: the Q3_K build Edge0/Edge0-145B-q3k (40.81 GiB) and the NVFP4 build Edge0/Edge0-145B-NVFP4 (47.79 GiB). Default to Q3_K: smaller file, less host RAM, faster.
Still unpublished. The official blog says the full evaluation setup, hardware environment and reproduction instructions will come with the technical report; as of Nov 20, 2026, the report is still unpublished, and AA, LiveBench and LMArena have no Edge0-145B listing.
Every fact on this page comes from the public sources below (the Hugging Face model cards and the official blog read directly on Nov 16, 2026; the GitHub release notes checked the same day).
More indexes: All Edge0 topics · All LLMs · All models