News · 2026-11-13

Edge0-145B's '29 GB' laptop headline holds only for GPU memory; the card measures about 61 GiB in total and names Qwen as the base

Edge0 released Edge0-145B on November 12, 2026, an open model with 145 billion total parameters and 10.12 billion active per token, under a blog headline promising "frontier AI capabilities on your own computer." The headline's "29GB peak memory" holds only for GPU memory. The model card's own measurement for the NVFP4 build is 23.25 GiB of VRAM plus 38.27 GiB of host RAM, about 61.5 GiB in total, and the card's metadata identifies Alibaba's Qwen3.8-Flash-Next as the base model.

Key facts

• Who and when: Edge0, November 12, 2026, in its launch blog post.
• Disk size: the card lists weights of 47.79 GiB for the NVFP4 build and 40.81 GiB for the Q3_K build.
• VRAM: the card measured 23.25 GiB of VRAM plus 38.27 GiB of RAM for NVFP4 on an RTX 5090 D with 24 GB, per the vendor; that is a vendor measurement, not an independent one.
• Independent runs: none found.

How it fits

Edge0-145B is a mixture-of-experts model: 48 layers, 512 experts per layer with 10 routed plus 1 shared active per token, a 256K context window and native 4-bit training. Its trick is a "latent sparse MoE." A shared down-projection takes each token to a narrower width, 1792 instead of 2560, and a shared up-projection brings the result back, so every expert is a small fixed record of 1,630,720 bytes. The engine keeps the hot experts on the GPU and streams the rest from host RAM; a 51.2-billion-parameter n-gram embedding table is memory-mapped from disk and read one row per token. A 1.26-billion-parameter draft head handles speculative decoding. (For the general ideas, see our lessons on mixture of experts and offloading and streaming weights.)

Headline versus fine print

The blog table says 29GB peak memory and its paragraph says 28GB. Further down it says about 40GB of experts sit in system RAM and about 48GB of vocabulary data on SSD. The model card's measured numbers are larger than the weights themselves because they include cache and buffers. The "40% of the memory" comparison pits Edge0's GPU-side peak against rivals' estimated total footprint, which the blog itself says "remain to be measured." The card's speeds, such as 79.6 tokens per second chat decode on an RTX 5090 D, are desktop-GPU figures. An Apple M2 Ultra row at 60.9 tokens per second carries a note that it "does not establish macOS release availability," and the build instructions target CUDA 13. HN commenters called the headline "a flat lie" and noted a 48GB MacBook would not cope with a roughly 70GB combined footprint.

Whose model is it?

The blog says "We train the model in its on-device architecture from the outset... avoids treating local deployment as a compression or pruning step." The Hugging Face metadata, however, carries base_model: Qwen/Qwen3.8-Flash-Next, a roughly 180-billion-parameter Alibaba model released in August. Edge0's own release-scope.json states "Model configuration identifies Qwen3.8-Flash-Next," sets actual_training_base_revision_verified to false, and adds "No training-source commit or independent ownership audit is asserted." It claims no ownership of Qwen materials. The two statements are not reconciled anywhere; heavy retraining after an architecture conversion is possible, but no evidence either way turned up.

Licence

The license (edge0-community-model-1.0, dated November 12) reserves three uses for a separate written license: selling, renting or licensing the weights; operating a model service for third parties financed by fees, subscriptions, advertising or sponsorship where customers control inference inputs; and offering an independent AI coding or office-productivity product that is financed the same way, hosted or on-device. Single-purpose domain tools and incidental assistant features are excluded, and community use including fine-tuning and sharing is allowed. The release-scope file also says it is not evidence of a separate Qwen commercial authorization. Our lesson on open-weight models explains why such licenses vary.

Benchmarks

All benchmark numbers come from Edge0's internal evaluations. It reports 64.6 on a coding-agent test against 58.7 for its Qwen base, 86.52 on a terminal test, and a six-test mean of 72.8 against 71.7 for Qwen. Edge0 trails Qwen on GPQA-Diamond (90.4 against 91.7), on an instruction-following test and on Humanity's Last Exam. A commenter called the Claude comparison, against Opus 4.6, outdated.

Reception and next steps

The Hacker News thread drew 281 points and 82 comments. The technical report with setup and reproduction instructions was not yet out. Until someone independent measures a laptop run, the story is the gap between a 29 GB headline and a measured 61.5 GiB total.

Primary source, verified: read the paper →

Ground Truth. (2026, November 13). Edge0-145B's '29 GB' laptop headline holds only for GPU memory; the card measures about 61 GiB in total and names Qwen as the base. Ground Truth. https://groundtruth.day/news/edge0-145b-headline-says-29-gb-the-model-card-measures-about-61-gib.html

@misc{groundtruth:edge0-145b-headline-says-29-gb-the-model-card-measures-about-61-gib,
  title  = {Edge0-145B's '29 GB' laptop headline holds only for GPU memory; the card measures about 61 GiB in total and names Qwen as the base},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/edge0-145b-headline-says-29-gb-the-model-card-measures-about-61-gib.html}
}

Topics: edge0 · open-weights · local-inference · mixture-of-experts · qwen · licensing