“Take work away from frontier LLMs”: Subquadratic’s decision model claims 12x Jev’s context window
Subquadratic launches SubQ Snap, a decision model with 384K-token context window — 12x Jev's 32K limit — claiming it can process 100K tokens and return a decision in ~1 second. Built on Subquadratic Sparse Attention (SSA) for linear attention scaling, Snap targets enterprise agent workflows: file selection, tool routing, test verification. Benchmarks show Snap beating Jev on 13 of 16 enterprise tests (Decision LCR 74 vs 48, Nimble 82.9% vs 76%), though Jev wins on When2Call. All results are self-reported; independent verification pending. Pricing at $0.05/M input tokens, free outputs. The real question: does a 384K window actually eliminate retrieval/chunking fragility for coding agents, or just push the bottleneck elsewhere?
Transcript
Masonry Subquadratic dropped Snap yesterday. Decision model. Three hundred eighty-four thousand token context window. That's twelve times Jev's thirty-two K.
Eyre Right.
Eyre And they're claiming a hundred K tokens in about a second. Five cents per million input, zero for output.
Masonry Which is the whole pricing story — it's a read-heavy model. Returns probabilities over allowed answers, doesn't generate text. Compute goes into the context, not the decode.
Eyre Mm-hm.
Eyre SSA — Subquadratic Sparse Attention. They introduced it in May. Linear attention scaling instead of quadratic. That's the architecture bet.
Masonry Yeah. Whedon's line: improving the underlying algorithms matters regardless of how much the big labs spend on compute. Otherwise we're leaving intelligence gains on the table.
Eyre Okay, that's genuinely a good line.
Masonry Right? The vacuum-of-vision quote is the other one. Decision models become exciting when they actually take work away from frontier LLMs.
Eyre Not just routing or evaluation. Actually doing the work the big model would've done.
Masonry Exactly. Coding agent fixing a bug across modules — error in one file, root cause in another, clues in tests and docs. Thirty-two K forces you to chunk or discard. Three eighty-four K… you feed the whole relevant slice.
Eyre Mm-hm.
Eyre But the benchmarks are all internal. Decision LCR seventy-four versus forty-eight. Nimble eighty-two-nine versus seventy-six. Thirteen of sixteen enterprise tests. Jev wins When2Call though — the 'should I invoke a tool' judgment.
Masonry Which is the judgment call, not the retrieval call. Interesting split.
Eyre And Snap's not on the latest Decision Index. Whedon says the benchmark authors told him v0.3.2 focuses on open-weight models with Jev as an exception, plus private tests now. So the published 0.2.1 score can't be directly compared.
Masonry Convenient. But they did release the long-context eval code. RULER adapted — ninety-six to ninety-nine point eight across thirty-two K to one million tokens.
Eyre One million tokens tested. But the public preview caps at three eighty-four K because… compute capacity. Two point six x more compute for the full million.
Masonry Right. Temporary restriction, they say.
Eyre Temporary until it's not.
Masonry Ha. Okay, that's fair.
Masonry The Workday thing from Tuesday — Jev picking UI components for Sana on the fly. That's the decision-model-in-production proof point. Snap's pitch: same workflow, but the agent sees the whole component library at once instead of a retrieved subset.
Eyre Mm-hm.
Eyre Whedon admits retrieval won't disappear. Enterprises have corpora way bigger than three eighty-four K. But the window's big enough that retrieval can be coarse — whole documents or modules instead of carefully trimmed snippets. Removes a lot of the fragile parts.
Masonry That's the product claim. Coarse retrieval equals fewer chunking decisions equals fewer places to silently drop the critical context.
Eyre Or it moves the failure mode to 'which three eighty-four K chunk do I feed it.'
Masonry Exactly. That's the bet.
Eyre Seventy-thirty someone reproduces the Decision LCR gap within a month and it holds within five points.
Masonry I'll take the other side. Sixty-forty the reproduction comes in lower — long-context evals are notoriously sensitive to prompt formatting and retrieval setup.
Eyre Fair. We'll know soon enough.
Masonry Twenty-nine million seed at five hundred million valuation. Formerly Aldea. Justin Dangel and Whedon — Whedon was Meta, then head of generative AI at Tribe. CEO swap a few months back.
Eyre Right. And they were already looking at attention-free architectures in July.
Masonry Snap's the sparse-attention play. Next one might drop attention entirely.
Eyre Okay. Snap public preview. Link in the show notes. We'll see if the context window actually ships better agents or just longer contexts.
Masonry That's the episode, Eyre.
Eyre Yeah. Good one.