AI Brief, 13 November 2026: the random numbers that name the GPU

The most substantial publication of the window is not a model. It is a forensic argument that the initial weights of six Chinese frontier models carry a signature of the GPU that generated them, and that the signature points to NVIDIA Blackwell parts barred from export to China. It comes from one researcher, it rests on a mechanism anyone with the weights can test, and no company named in it has responded.

• Checkpoint forensics. Will DePue reports that DeepSeek, Moonshot, MiniMax and Ant Group models read as 148- or 152-SM hardware, against 132 for Hopper. No code published.
• Gemini 4 Hydrogen, 12 November, USD 1.50 and 7.50 per million tokens. Artificial Analysis independently scores it 54, and ranks it 16th of 78.
• Inkling 2 Small's weights landed at 19:53 UTC on 12 November under Apache 2.0, closing a promise tracked here since the 6th. Running them takes at least 500 GB of aggregate VRAM.
• Edge0-145B claims frontier work on a laptop. Its own licence file concedes the weights are a Qwen derivative, the download is 105 GB, and both repositories show zero downloads.

A fingerprint left by 2,048 threads per core #

Will DePue, writing at Unemployment Labs on 12 November, keeps a narrow claim and a broad inference separate. The narrow claim is that six models' published weights were initialised on Blackwell-class GPUs. The broad inference, which he says the evidence "supports, without proving," is that Blackwell also did the training.

The mechanism is a detail of how PyTorch fills a tensor with random numbers. Initialisation uses the Philox counter-based generator through cuRAND, which is deterministic: the same seed yields the same stream. To fill a large tensor quickly, cuRAND divides the work across GPU threads, and for a large enough allocation it uses the hardware maximum of 2,048 threads per streaming multiprocessor. Thread count therefore falls out of the chip's SM count, which differs between generations:

Hardware	SMs	Threads
H100 SXM, H200, H800 (Hopper)	132	270,336
B200, B300 (Blackwell)	148	303,104
GB200, GB300 (Blackwell)	152	311,296

Each thread writes every $T$ -th position, where $T$ is the total thread count, so the value at flat index $i$ comes from thread $imodT$ at offset $\lfloor i/T\rfloor$ in that thread's substream. Change $T$ and the same numbers land elsewhere, so the thread count is recoverable from the weights: it is baked into where each number sits.

The trace survives training because some weights are never trained. Embedding rows for reserved vocabulary tokens that appear in no document receive no gradient, so they still hold their initialised values almost exactly. DePue searches roughly six trillion seed-and-layout combinations for the one regenerating those values, then checks the winner against held-out weights.

That bounds what the method can see. Take DePue's figure of 745,472 attention weights per layer. Below index 303,104 both Blackwell layouts assign thread $i$ at offset 0, so the predictions are identical and the region is silent. Divergence starts there, 40.7% of the way through the layer: at index 400,000 the 152-SM layout reads thread 88,704 and the 148-SM layout thread 96,896, unrelated numbers. Separating B200 from GB200 therefore rests on about 59% of each layer, and a tensor under 303,104 elements cannot do it at all.

The dependence is visible in three lines:

import torch
sms = torch.cuda.get_device_properties(0).multi_processor_count
print(sms, sms * 2048)   # 132 270336 Hopper | 148 303104 B200/B300 | 152 311296 GB200/GB300

The results are lopsided. DeepSeek-V4-Flash reads as 148 SMs at 1,385σ with an exact copy of reserved rows; Kimi K3 and K3.1-Pro both read 152 SMs from the same seed at 305σ; MiniMax-M3 reads 148 at 323σ; Ant Group's Ling-3.1-flash shows untrained rows correlating at $r=0.98$ ; DeepSeek-V4-Pro is weakest at 49σ. DeepSeek-V3, Kimi K2.5, MiniMax-M2.7 and Ling-3.0-flash all read as Hopper, and DePue puts the chance threshold for a search that size at about 8σ.

The controls are the strongest part. On a physical B300, unmodified PyTorch reproduced 461 of 469 selected DeepSeek-V4-Flash rows, 445 bit for bit; on an H100, all 57,344 preserved DeepSeek-V3 values bit for bit; on a rented GB200, all 13 preserved Kimi K3 rows. The method agreed with all 24 cases where a developer had disclosed its hardware, and NVIDIA's own Nemotron-3 Ultra reads as 152 SMs, matching its disclosed GB200 training. Alibaba's Qwen3, initialised on CPU, leaves no fingerprint, the negative control the argument needs.

Three weaknesses deserve equal billing. No code or data is published — reproduction scripts are "available to reviewers on request," the one thing standing between this and independent confirmation. An AMD MI300X can be configured to present 148 or 152 units and produce the same layouts; DePue excludes this for DeepSeek-V4-Flash, where 363 of 377 differing values match NVIDIA and none match AMD, but did not test the other five. And the compute estimates are soft: about 12,000 Blackwell GPUs on a 7,000 to 22,000 range, roughly $850 million, which he says he is "not confident in."

What the paper does not claim matters as much. It does not establish where the chips sit, how they were obtained, or that any law was broken, and initialisation hardware need not be training hardware. Techmeme picked it up the same day; DePue says senior researchers at OpenAI and Anthropic verified the results, which cannot be confirmed from anything published. Nvidia, DeepSeek, Moonshot, MiniMax and Ant Group have all said nothing.

Gemini 4 Hydrogen, and what an index of 54 is worth #

Google published Gemini 4 Hydrogen on 12 November, reaching OpenRouter at 18:01 UTC. It is the cheap, fast member of the Gemini 4 family rather than a new frontier tier: $1.50 per million input tokens and $7.50 per million output, against $2.00 and $10.00 for Carbon and Argon, with a 1,048,576-token context and a batch tier at half price, and a new xhigh thinking level above low, medium and high.

The headline figure is checkable, and it checks out. Google reports an Artificial Analysis Intelligence Index of 51 at high and 54 at xhigh; Artificial Analysis independently lists 54 at xhigh, with 188 output tokens per second at the same price. It also ranks the model 16th of 78, the fact Google's post leaves out, since the comparison it draws is to Gemini 4 Argon at 50, inside its own family.

The rest of the table mixes third-party boards with Google's own. Terminal-Bench 4.1 at 57% against 18.2% for Gemini 3.8 Flash is a large jump on a public benchmark. Against that, the Multimodal Knowledge Agent benchmark is Google's own, the 96.7% honesty rate and 99.9% chain-of-thought legibility over 200,000 audited turns are internal, and the safety deltas — overreach down "up to 81%", misaligned actions "cut 12x" — are unaudited and relative to an unnamed baseline. The post cites Terminal-Bench 4.1 for the model and 4.0 for Antigravity's Orchestrator, and its image caption says Gemini 3.5 Flash where the text says 3.8.

The 30 November price change flagged here on the 11th applies to Carbon, not Hydrogen.

Where the weights actually live #

Three releases in this window answer the same question — can a large model run on hardware someone owns — with very different candour.

Thinking Machines published Inkling 2 Small under Apache 2.0, with the weights uploaded at 19:53 UTC on 12 November. This closes a promise flagged here in six consecutive briefs since the 6th, when the model was callable but unreleased and nothing had appeared under thinkingmachines since July. The lab's own wording was the week of 9 to 15 November, and it landed inside that window.

It is a 48-layer mixture-of-experts model routing each token to 8 of 384 experts, natively multimodal, with a 1M-token context, pretrained in NVFP4. The card says 594B total and 17B active while Hugging Face's metadata says 383B, a gap the card does not explain and which is arithmetic rather than a contradiction: 247.7B of those parameters sit in a U8 block, and NVFP4 packs two values per byte.

The figure that matters more is the hardware line: at least 500 GB of aggregate VRAM for NVFP4, two B300s or four H200s, and 800 GB for MXFP8. Nine hours after publication the repository showed 126 likes and 12 downloads. Apache-2.0 weights needing a multi-GPU server are a real contribution to the research commons, not a local model.

Edge0-145B, published 12 November, claims the opposite: "frontier AI capabilities on your own computer," 145B total with about 10B active, 80+ tokens a second of decode on a 48 GB laptop at about 120 W. It drew 127 points on Hacker News.

The metadata tells a different story. Both Edge0-145B checkpoints carry base_model: Qwen/Qwen3.8-Flash-Next and base_model:quantized, making this a derivative of Alibaba's 180B model rather than a new one. The blog calls it open-source; the licence is Edge0 Community Model License 1.0, dated the day of release, and its own text is admirably clear that it is not a weights licence: "Upstream Materials means Qwen materials … They are not relicensed as Edge0 property." The 29 GB peak is real, but it is a resident working set, not a download: q3k is 105 GB across 93 files.

The benchmark table is self-reported, with a technical report promised but not published, and its shape is revealing. Edge0-145B's mean across six benchmarks is 72.8 against 71.7 for the base model it is built from, and all of that advantage comes from two rows: DeepSWE 1.1 (64.6 against 58.7) and Terminal-Bench 2.1 (86.52 against 83.3), contributing +1.52 to the mean. The other four subtract 0.38 — GPQA-Diamond 90.4 against 91.7, IFBench 80.6 against 81.3, HLE 35.4 against 35.9. Gains concentrated in two agentic benchmarks with small losses on knowledge and instruction-following is the signature of post-training on agentic traces, not of a quantisation recipe. Neither repository had been downloaded once, two days after publication.

smolmaxxing is the honest version of the same ambition, and drew 300 points on Hacker News: an MIT-licensed patch set on llama.cpp with 335 stars and a notebook recording 160 experiments. On a Xiaomi TV box with about 1 GB of usable RAM it runs Qwen3.5-2B at 7.7 tokens a second and a 16B mixture-of-experts model at 5.1. On a 2013 iPad Air with roughly 700 MB usable it gets 24 to 26 tokens a second from a 1-bit 1.7B model and 2.6 to 4.5 from the same 16B. The techniques are specific: a 1-bit lookup-table kernel needing only basic NEON for the A7, which has no dot-product instruction, and expert paging from flash with next-layer prefetch.

Its most instructive number is one it did not headline. On the iPad, an 8B dense model streamed from flash manages 0.08 tokens a second; an 8B mixture-of-experts model manages 3.5, roughly 44 times faster at the same parameter count. When weights live on storage rather than in memory, sparsity stops being a compute optimisation and becomes the difference between working and not. That is the bet Edge0 is making, measured by someone with nothing to sell.

Also notable #

• Jensen Huang resigned as Nvidia's chief executive, the loudest item of the window at 3,152 points on Hacker News. The 8-K dates the board's acceptance to 11 November, effective immediately, across his roles as president, CEO and director. CFO Colette Kress is interim chief executive pending a search; Harvey C. Jones, a director since 1993, becomes the company's first board chair. The board had earlier opened an investigation into reports Huang was targeted by a foreign intelligence effort, finding "no reason to believe that either national security or NVIDIA secrets were compromised."
• Anthropic's IPO prospectus went live, listing Class A stock on Nasdaq under ANT, pricing expected 18 November and trading 19 November. Disclosed Q3 2026 revenue is $17.0B, more than 12 times year on year, against a net loss of $23.3B that includes a $21.9B non-cash charge from the valuation increase.
• Huawei's Zurich lab published MACaQ under Apache 2.0, a model-agnostic quantiser for GGUF, vLLM and Diffusers, with arXiv:2611.10649 behind it and 215 stars. It quantises a 122B model in 64 minutes using 32.5 GB of device memory, and reports W4A4 perplexity of 2.8994 on Qwen3.8-27B against 2.8463 at BF16, with prefill 2.33x and decode 2.58x faster. The figures are the authors' own. Two arXiv papers extend the same thread, neither independently evaluated: SteerDistill (2611.13410) and Precision on Demand (2611.14418), which raises weight bit-width only at uncertain decoding steps.

What to watch #

• Whether anyone reruns the checkpoint forensics on weights DePue did not test. This is the rare claim whose falsifier is cheap and public: the weights are open, PyTorch's initialisation path is public, and the prediction is specific enough to break. The decisive step is publishing the search code, currently held back for reviewers; until then the strongest result in the window rests on one analysis pass. A named lab could settle its own case in an afternoon by publishing its initialisation seed and thread count, and Qwen's CPU-initialised checkpoints show there is nothing shameful in that answer.
• Whether Inkling 2 Small gets a score from anyone who did not train it. Apache-2.0 weights with a roughly 60-benchmark self-reported table and 12 downloads is the usual gap between a release and a measurement. The figure worth more than any benchmark row is measured memory at inference against the card's 500 GB floor, which decides whether the licence means anything in practice.
• Whether Edge0 publishes the technical report and a licence covering the weights. The thing to look for is an ablation separating quantisation from post-training, because the published table credits a compression recipe with a gain that sits entirely in two agentic benchmarks.
• Still open: Voxtral 3 weights, flagged daily since 21 October, with nothing under mistralai since 15 October. No independent number for Tencent's DSpine drafter, Nanbeige4.3-33B or Ripple-65B Preview; no placebo control for the weight-space programming result; no assessment from the three complexity theorists who agreed on 26 October to examine the Exponential Time Hypothesis manuscript; and no fail-open rate from any decision-model vendor, asked here since 10 October and now 35 days.