It Renamed DeepSeek's Architecture. Then the Weights Told the Truth.

On July 28, 2026, SK Telecom — the Korean telco — released A.X K2, a 688-billion-parameter Mixture-of-Experts model, positioned as a from-scratch foundation model for Korea’s Sovereign-AI effort. The model card reads like a sheet of original engineering: “Sparse Gated Attention (SGA),” “Gated Norm,” “Gated Transformer Blocks,” all “trained from scratch.”

Then you read the config, and the names dissolve.

The tell

A.X K2 declares a brand-new model_type: axk2 — and ships no modeling code to go with it. That’s the first thing a scanner notices: a novel type name with nothing novel behind it. Look at the actual tensors and the architecture is unmistakable:

• Multi-head Latent Attention (MLA) — DeepSeek’s low-rank attention. (The card does name this one.)
• A sparse-attention “indexer” with exactly DeepSeek-V3.2’s lightning-indexer tensors (indexer.wk, indexer.wq_b, indexer.weights_proj). The card calls it “SGA.” It is DSA — DeepSeek Sparse Attention.
• DeepSeek’s MoE, bolt for bolt: 256 routed experts + 1 shared, top-8, sigmoid noaux_tcrouting, one dense first layer. Hidden 7168, 61 layers, FP8. The DeepSeek-V3.2 signature.

90,721 of A.X K2’s tensors carry DeepSeek-V3.2’s exact names. At this point it would be easy to publish the word “relabel” and move on.

We didn’t — because matching names and shapes are exactly what you’d see whether A.X K2 copied DeepSeek or merely adopted its design. A config cannot tell those two apart, and they could not be more different.

Adoption vs. copy — and why only weights decide

Two very different stories produce an identical config:

1. Copy. Take DeepSeek-V3.2’s weights, rename the type, ship it as your own. That makes “trained from scratch” false.
2. Adoption. Build the DeepSeek-V3.2 architecture and train your own weights into it from zero. That makes “trained from scratch” true — you adopted a design, not a checkpoint.

The tensors’ names and shapes match either way. The only thing that separates a copied checkpoint from an independently-trained one is the numbers inside the tensors — the weights themselves. So we diffed them.

The weight diff

We compared A.X K2’s weights against deepseek-ai/DeepSeek-V3.2, tensor for tensor, on the layers they share — 768 experts plus the attention and the indexer — dequantizing the FP8 on both sides so the comparison is apples to apples.

For scale: two independently trained weight sets of similar magnitude differ by roughly 140% here (they’re essentially uncorrelated); a genuine copy reads near 0%. (We compared the shared layers only — the embeddings differ trivially because A.X K2 ships its own tokenizer, so they’d prove nothing either way.)

So the verifiable half of the claim is settled: A.X K2’s weights are independent of DeepSeek-V3.2’s. Whatever they are, they are not DeepSeek’s checkpoint — a copy and a light fine-tune are both ruled out. A weight diff can prove not-a-copy cleanly; it can’t by itself separate “from literal zero” from “heavily re-trained off some other seed” — but nothing here points to an inherited checkpoint of any kind.

The honest verdict

A.X K2 is not a re-badged DeepSeek checkpoint. On the evidence it is an independently-trained 688B model — with real additions of its own (a Gated Norm the base architecture doesn’t have, a head-specific attention gate, and SK Telecom’s own tokenizer and data). It is, however, a model built on DeepSeek-V3.2’s architecture — MLA, DSA, and the DeepSeek MoE recipe — with its own weights.

That makes A.X K2 notable: the first non-Chinese adopter of DSA (DeepSeek Sparse Attention) we’ve weight-verified. As of this writing (late July 2026), DSA had — in everything we’ve scanned — shipped only from a handful of Chinese labs. We won’t claim it’s the first non-Chinese adopter anywhere — that would take an exhaustive global audit we haven’t done — but it’s the first we’ve confirmed at the weight level, and Korea belongs in that conversation now.

The one thing the card gets wrong isn’t the weights — it’s the attribution. Renaming DSA to “SGA,” presenting DeepSeek-V3.2’s block as an SK Telecom “Gated Transformer Block,” and listing no base architecture leaves a reader thinking the design is original when the design is DeepSeek’s. That’s an attribution gap, not a fabrication — and it’s worth naming precisely, because the difference between “they copied a model” and “they adopted an architecture and didn’t cite it” is enormous, and only the weights let you say which.

Why this is the whole point

A config-only scanner reads model_type: axk2 and either shrugs or, worse, sees the DeepSeek signature and cries “copy.” Both are wrong. The truth — independent weights on an adopted, uncredited architecture — lives in the tensors, and you only get it by reading them. Telling adoption from theft is exactly the line that matters for provenance, and it’s exactly the line a config can’t draw.

That’s what ModelDNA does: it doesn’t stop at the name on the box, and it doesn’t jump to the accusation. It reads the weights, and it says only what the weights support.

A.X K2 architecture facts are from its public config.json and safetensors headers; the weight comparison is a dequantized (FP8 block-scale) per-tensor Frobenius diff of the shared layers against DeepSeek-V3.2 — a measurement anyone can reproduce from the two public checkpoints. “SK Telecom” here is the Korean telecommunications company and its A.X model line — a legitimate frontier lab with no provenance history at issue; this piece concerns architecture attribution only.