A small model that punches well above its weight.
Big models are brilliant consultants who have never seen your files. We build the other thing: a small model that knows your material, runs inside your walls, and doesn't need the expensive hardware to do it.
The trick isn't making it smarter. It's giving it a filing cabinet, so the part that thinks stops wasting itself re-deriving what it could have looked up. The fastest-growing piece of a frontier model is now exactly that cabinet — and it doesn't have to sit on the graphics card.
Read the bet See what's running“Don't you need a trillion parameters?”
Usually not — and often you're not allowed to use one anyway. The real choice most companies face isn't a small model versus a huge one. It's a model that can read their data versus no model on that data at all.
Size isn't the axis
What grows with a bigger company isn't how hard the thinking is — it's how much of their own material the system has to know. Part numbers, procedures, contract terms. That's looking things up, and looking things up is cheap.
Where the data may go
The frontier model you're picturing is frequently one your legal department won't let you point at your files. A capable model inside your own walls beats a better one you can't use.
The big labs are already moving
Alibaba built 28% of a brand-new flagship as a lookup table on purpose, so it wouldn't need the expensive memory — and said so in the release notes. This isn't a small-model excuse. It's where the field is heading.
When you do want the big one
Genuinely novel reasoning, long-horizon planning, unfamiliar code at scale. A filing cabinet adds knowledge, not thinking. When that's what you need, we'll say so.
Parameters that don't have to sit in fast memory
Qwen3.8-Flash-Next is 125B parameters. 51B of them — 28% — are not attention, not experts, not a network. They're a phrase book: keyed on the last two or three tokens, learning nothing while it runs, read sixteen rows at a time. The lab that built it said plainly why.
“Embeddings are more amenable to offloading than mixture-of-experts. Indexing with short engrams makes parameter scaling highly efficient for memory constrained accelerators.” — Qwen3.8-Flash-Next model card
They grew the model along the one axis that doesn't have to sit on the accelerator, and said so in the release notes. Three labs arrived there independently — Google with per-layer embeddings in Gemma 3n, DeepSeek with Engram, Qwen third. Convergence by three groups is better evidence than any one of them being clever. The technique is public. Only the weights are encumbered.
What follows is a change of question. Not how many parameters do you have, but how few of them have to sit in the fast memory. Said honestly, and it matters that it's said honestly: nothing in this architecture makes memory cheaper. It makes slow memory usable. Those aren't the same sentence.
The bet, stated so it can lose
A phrase table may not need to be trained at all.
Qwen trained theirs jointly with the model. But an engram table is keyed on the previous few tokens — the address is knowable before any computation happens, which is why it can sit at layer 2 and be fetched while layer 1 is still busy. A structure with a knowable key, no runtime learning, and no gradient path at inference is closer to an index than to a network.
- Hypothesis
- The table can be seeded from corpus n-gram statistics rather than learned end-to-end, with only the gating and projection tuned afterwards.
- If true
- The largest single component of the architecture becomes near-free to construct, and building a model in this shape stops requiring a frontier lab.
- If false
- We learn what the table stores that counting can't capture — which is the more interesting result of the two.
- The test
- One small teacher, two students: table built by counting versus table learned jointly, held against the same evaluation. Weeks and credits, not years and millions.
- Provenance
- Distilled from Laguna (Poolside) — a US lab that trains from scratch, so there's no foreign base model underneath. XS.2 is Apache-2.0; S 2.1 is open weights under OpenMDW-1.1. OLMo is the alternate when a buyer's question is “what's in the corpus” rather than “who built it.” Deliberately not from Qwen: that would make the result a derivative and re-attach the commercial license we'd otherwise never touch.
What already runs here
We are not tied to a model. AI research and software development is the work; the particular models are this quarter's instruments and we expect to replace them. What's durable is being able to walk into a shop and make whatever they have run inside whatever constraints they have — whichever model, whichever card, wherever the data isn't allowed to go. On one box here: half a dozen different architectures and quantizations, across GPU and CPU lanes, swapping when a provider dies.
Not a demo spun up for a screenshot. The instrument runs continuously, fails in observable ways, and has its failures written down — which is why every design decision below traces to a specific logged one.
Continuous multi-agent systemRUNNING
Heterogeneous model seats, local CPU and GPU inference lanes, a message relay, and scheduled work — up and coordinating, not launched per demo.
External memory the agents depend onRUNNING
One archive in three indexed forms, supersede-never-delete semantics, write-ahead log. It exists because two instances of the same agent on two machines grew different memories, and continuity had to become a file rather than a hope.
Ternary discipline: +1 / 0 / −1RUNNING
Zero is held-open possibility, not absence. Unresolved things stay unresolved instead of collapsing to a guess. Our ternary work to date is weight-format mathematics — it has not been evaluated against a model.
Function tests, not existence testsRUNNING
Every artifact is checked for what it does, never for whether it's there. The rule was written after a document with the shape of an official record turned out to have been typed off a screen — and it caught the same failure again, better disguised, a day later.
Hardware: one box — AMD Ryzen 9 9950X3D, 96 GB DDR5-6000, Radeon RX 7900 XTX. The configuration the thesis is about: one accelerator next to a lot of ordinary memory.
The studio is the stress test
A system that only runs in benchmarks isn't running. Four animated channels put the fleet under real production load every week — scripts, multi-voice narration, art direction, release scheduling — which is how the failure modes above got found in the first place. The output is real. It is evidence, not the product.
Conan & the pulp sagas
Savage tales told whole, in Frazetta oil-paint fury.
Lovecraft & the dread
Cosmic horror in Giger-meets-Mœbius biomech surrealism.
Heroes & hidden history
The heroes the record skipped, told with receipts on screen.
Tales that teach
Classic fables in watercolor storybook style.
Murasaki — an always-on synthetic host, rigged and lip-synced, with a wardrobe systemIN DEVELOPMENT Her voice loop and continuous show are being built; neither is live yet, and this page will say so until they are.
Where every piece of this came from
House rule: cite where the idea came from, then cite where they got it, as far back as it can be traced. Most of what's on this page is other people's work, and it should be legible which parts.
- The lookup-table idea reached us through Qwen3.8-Flash-Next's model card. Qwen say plainly they aren't first: Google did per-layer embeddings in Gemma 3n, DeepSeek published and open-sourced Engram in January, Qwen arrived third and independently. That convergence is the evidence, not any one of them being clever.
- The offload measurements are not ours. 84→60 GB per card with bit-identical output was the SGLang team. Three builds spanning 25 GB on disk that all load at 63 GB was a third party publishing quantizations on Hugging Face. We have measured none of it and won't imply otherwise.
- The hybrid backbone is Ornith-1.5-9B (ornith-ai, MIT) — state-space layers with full attention every fourth, descending from Qwen's hybrid design.
- Dynamic quantization by layer sensitivity is Unsloth's, built on the importance-matrix approach.
- The open-weight US-provenance teacher is Poolside's Laguna.
What's actually ours
Three things, and no more than three: the hypothesis that such a table can be counted rather than trained; using the shift in a model's importance matrix as a cheap early read on whether the table is absorbing real work; and the argument that this is what makes serious tooling affordable for operators nobody else serves.
- +1earned — proven, with a falsifier that could have failed
- 0held open — possible, not yet earned, never evicted
- −1a scar — known wrong, and kept
What we are not claiming
A page like this usually inflates. This one lists what it can't support, because a claim you can check is worth more than one you can't.
- No revenue, no customers, no employees.
- No benchmark results of our own. None have been run.
- No claim that ternary weight formats have been evaluated on a model. They have not.
- No claim to have reproduced Qwen's architecture. We have read their report.
- The engram-by-counting hypothesis is a hypothesis. It has not been tested.
Everything on this page that isn't in that list is either running today or stated as a plan.
Work with us
Deployed agent stacks and systems work on your infrastructure — your data stays home. Research collaboration and compute partnerships welcome; the experiment above is bounded and specific, and we'd rather run it with someone than alone.