Reading models at the tensor level.
Architecture intelligence from ModelDNA — what config files, weight layouts, and modeling code reveal about frontier models, before anyone downloads a byte.
What It Takes to Build a 4B Model From Scratch
A modern 4B language model is trained on about 25 trillion tokens: at least 6 × 10²³ operations and around 420,000 H100 hours for the final run, checked against a real model card. Where the numbers come from, a simple modern design shared by Qwen3, SmolLM3, Granite 4, OLMo 2 and Nanbeige 4, and what a pilot, a first build and a stretch budget buy.
Small OCR Specialists, Tested on Long Documents
A 0.85B-active document model beat a 120B general vision model by 34 points on 1,767 pages. Two single-page rivals from LightOn matched it on plain text and fell behind wherever text crossed a page break. What each model does and doesn’t do, what we tested, and which to use for which job. A living post, updated as models and tests are released.
Where Your Data Goes: Open Weights, Hosted APIs, and the Risk You Can Actually Control
A model file can’t send your data anywhere. A service can. Whether a model’s developer can see your data depends on how you deploy it, not on who made it. Choose the model on measured performance, then contain the risk with architecture: self-hosting, no outbound network, vetted serving software, and output review.
The Prompt, Not the Passport
We retested the claim that Chinese coding models write weaker code for U.S. customers: four Chinese and two U.S. models, 56,160 pre-registered, exploit-checked responses. A U.S.-government label never made code less secure. Describing the stakes made every model more careful, whichever country the prompt named.
We Checked DeepSeek’s New Model Card on Our Own GPUs
Six open models, four public benchmarks, each at its own card’s settings. DeepSeek-V4.1’s automation and factual-recall gains reproduced; its claimed math gain did not. A Xiaomi model matched it at automation on half the GPUs and a third of the energy — and the strongest models almost never said “I don’t know.”
The Confidence Is the Product
A new class of model answers in one forward pass and tells you how sure it is — and for a guardrail, that confidence number is the product. We measured four of them on two domains none had seen. Calibration error ranged from 0.081 to 0.599, the ranking did not survive the change of domain, and the worst performer stayed confident on half the cases where we had deleted the evidence it needed.
Importing the Vocabulary
DeepSeek published a new attention mechanism last September; six labs shipped it within a year and every one is Chinese. That was the finding we expected. The one we did not: six American labs are already building Chinese techniques into architectures of their own — and the deepest multi-token prediction implementation in our atlas turns out to be American.
You Wouldn't Hire by Skull Size
A benchmark score is not a measurement of a model — it's a measurement of a model plus a harness, and almost nobody reports the second half. What a harness actually contains, why the grader is the part everyone skips (including us: two of our own checkers were handing out points), and the disclosure card we now publish beside every score.
Two Things We Got Wrong About Abliterated Weights
We claimed a censorship-removal edit couldn't be extracted from mixture-of-experts weights. To test it properly we designed an experiment that could refute us — and it did, twice. What survives is a cleaner distinction: whether an edit can be attached is a storage property; whether it can be recovered is a property of the run that made it.
A Model Can Be Aligned and Uncensored at the Same Time
Routed-adapter models — a base plus swappable LoRA specialists and a router — can make 'alignment' a per-request decision that a static weight scan can't see. We built the check, ran it on a shipping routed model (clean on our probe), then proved the check has teeth by recovering an abliteration as a LoRA and catching it: refusal 100% → 47.5%. The threat turns out to be architecture-dependent.
You Can Ablate a Refusal. You Can't Ablate a Belief.
The 'Chinese models sabotage American code' risk didn't appear when we measured it — so where's the real, measurable state-alignment signal? In the worldview of the text. A Chinese-lineage model refuses some topics, states the official line on others, and keeps that line even after we surgically remove its ability to refuse. Strongest in Chinese. One model, scoped, reproducible.
A Threat Scenario Is a Hypothesis. So We Tested One.
A report raised a risk — that Chinese models might write worse code for the U.S. government. It's a threat scenario, not a fact, and the right response is to test it, not ban on it. We ran a preregistered experiment on a first-party Chinese model, with confidence intervals, and found no evidence of the effect — only a benign stakes-framing effect. Full reproducible method, so anyone can add evidence.
Do Chinese AI Models Write Worse Code for America? Here’s How You’d Actually Find Out
A widely-shared report says Chinese code models degrade under U.S.-government framing and hedges on whether it’s deliberate. The refusal findings hold up; the “on purpose” claim is a confound. The fix isn’t a ban — it’s to measure the code directly under a compliance-matched control, use abliteration to probe the cause, and verify-then-contain.
CISA’s Open-Weight AI Guidance Points the Test at the Wrong Artifact
New federal guidance tells agencies to treat any AI model without full training-data access as “proprietary.” But the inspectable artifact for a model is the weights, not the data — and the one paper CISA cites argues the other way. The rule discounts the most inspectable option. The fix: verify the weights, then contain the rest.
It Renamed DeepSeek's Architecture. Then the Weights Told the Truth.
SK Telecom's A.X K2 ships a novel-sounding architecture and a 'trained from scratch' claim — but the config is DeepSeek-V3.2 (they renamed DSA to 'SGA'). A per-tensor weight diff (median 113% Δ) shows the weights are genuinely its own: adoption, not a copy. The line only a weight diff can draw.
Who Actually Adopted MLA? We Counted.
We counted every fine-grained MoE model in the atlas. Original US adoption of MLA: zero (the 20% was all NVIDIA re-quants of Chinese models). MLA is a five-lab club — and even China isn't unanimous. A data-backed adoption survey.
What Solar-Open2 Reveals About the Frontier
A 250B model isn't chasing the frontier — which is exactly why its architecture is such a clean signal. Solar-Open2's choices reveal which frontier ideas have become table stakes (the MoE recipe) and which are still risky bets (MLA). Architecture as a revealed preference.
The Trick That Makes Kimi-K3’s 2.78 Trillion Parameters Almost Free
Everyone’s counting Kimi-K3’s parameters. We found the move that makes them cheap — a “latent MoE” that runs the experts in a compressed space — reverse-engineered from public metadata without downloading the model.