Subquadratic AI is a Miami startup with a $29M seed round, a closed-weights one-million-token LLM called SubQ, and a benchmark deck that has the rest of the field arguing. This note collects the facts, the architecture, and the honest reasons to be skeptical, in one place. If you came here from search for "subquadratic stock," skip to the stock question.
The company in three sentences
Subquadratic AI emerged from stealth on May 5, 2026 with a $29M seed at a reported $500M valuation. The company is based in Miami and led by a credentialed CEO and CTO with prior work at large research labs. It has filed an SEC Form D for the seed, has an active hiring pipeline, and disclosed a $19.6M, 24-month GPU rental contract with the NASDAQ-listed Digi Power X.
What SubQ actually is
SubQ is the name of the model. The production preview, SubQ 1M-Preview, is a closed-weights long-context LLM with a one-million-token context window served over an API. A research configuration extends to twelve million tokens, with fifty million tokens targeted for Q4 2026. The architecture is called Subquadratic Selective Attention (SSA), and the selling point is that attention complexity grows slower than the standard O(n²) bound for long sequences.
The mechanism, as described in Subquadratic's technical report:
- Each query token does not attend to every prior key.
- A learned selector picks a small set of keys for each query.
- The selector itself runs in subquadratic time, not just the attention that follows it. This is the technical hinge: prior "sparse attention" methods (top-k, sliding window) often sneak a quadratic step into the selection.
- Attention is then computed only over the selected keys plus a local window.
It is the same family of ideas as Mamba (state-space sequence models), RWKV (linear recurrence), Kimi Linear, and DeepSeek Sparse Attention. The architectural debate is well-trodden, which is why SubQ's specific claim that both the selection step and the attention step are subquadratic is the part to argue about.
The benchmark situation
The production model's reported numbers are RULER-128K of 95.0 and MRCR v2 of 65.9. A single unnamed third party confirmed both. That is the strongest, most reproducible thing on the launch page.
Then it gets harder. The research configuration claims:
- MRCR v2 of 83.0 at the twelve-million-token configuration.
- A thousand-fold compute reduction at twelve-million-token context, relative to a quadratic baseline.
- 92.1 percent needle-in-a-haystack accuracy at twelve million tokens.
Those numbers are self-reported, run once, on a closed model with no public weights and no arXiv preprint. The technical report concedes the single-run point directly. None of this is a fraud signal by itself; it is the normal state of a young lab. It is also the reason an independent reproduction by anyone with API access would be more interesting than the launch deck.
Is there a Subquadratic stock?
No. As of May 2026 Subquadratic AI is a privately held startup. It has filed an SEC Form D, which is the standard private-placement notice, not a public-offering registration. There is no ticker symbol and no way to buy shares as a retail investor. The "subquadratic stock" search query is almost certainly people checking, just in case.
Why the AI community is skeptical
Three reasons, all separable from "is the team credentialed." The team is fine.
One. Prior subquadratic-attention claims at frontier scale have a mixed track record. Mamba and RWKV are truly subquadratic but underperform Transformer baselines at the largest scales. Kimi Linear is subquadratic in some configurations and quadratic in others. DeepSeek Sparse Attention is genuinely sparse but not strictly subquadratic in the selection step. The base rate for "we have subquadratic attention that matches quadratic at scale" is currently low.
Two. Magic.dev made a structurally identical pitch in August 2024: a hundred-million-token context model, roughly thousand-fold efficiency, over five hundred million dollars in funding. Twenty-one months later, no externally validated frontier model has shipped from that effort. The pattern matches enough to require some discount.
Three. The benchmark report explicitly states every number was run once. Frontier LLM evaluation has a known reproducibility floor that is wider than people admit; single-run benchmarks at the extreme end of context length are the wrong place to start.
The honest version
Subquadratic is a real, well-funded, recently launched closed-weights AI startup whose benchmark claims are aggressive and not yet independently reproduced. It is not Theranos. It is also not OpenAI or Anthropic. It is closer to Magic.dev: a real company with a real team and a launch deck whose extreme numbers should be treated as marketing pending third-party verification on a public API endpoint or an arXiv paper.
The full interactive breakdown lives at the Subquadratic lab on this site, with the SEC filing, GPU contract, benchmark table, comparison to DeepSeek and Kimi, and the “is this Theranos” question argued in both directions.
Where this fits in the long-context LLM map
SubQ is one of four long-context approaches currently in the conversation. The companion DeepSeek-V4 architecture lab covers the open-weights side of the same problem (CSA, Muon, FP4 KV cache). The LeWorldModel lab covers the JEPA-style world-model angle. The four-way map is the most useful frame for reading any of these launches.