Seven years of outcome-labeled social data just met a frontier-class open model we're allowed to own. This is the case for turning LunarCrush's dataset into LunarCrush's model, and the IP stack it builds.
For years, "train your own model" meant one of two bad options: spend tens of millions pretraining, or fine-tune a weak open model that embarrassed you next to GPT. July 2026 quietly ended that.
Thinking Machines shipped Inkling: a 975B-parameter mixture-of-experts model (41B active: big brain, small inference bill), natively multimodal across text, images, and audio, with a 1M-token context window. It's deliberately positioned not as a benchmark king but as the best base to customize. And it's Apache 2.0: we can fine-tune it, keep the resulting weights closed, deploy anywhere, and never owe anyone a royalty or a relationship.
Just as important: Tinker, their managed fine-tuning platform, means we don't buy GPUs or hire an ML team. Experiments run in the hundreds-to-low-thousands of dollars. There's a 50% discount window open right now. And a 276B "Small" variant (12B active) exists specifically for the high-volume, low-cost inference work that looks exactly like our firehose.
Everyone has posts; AI labs have trillions of tokens of them. Everyone has price history; it's public. What nobody else holds, and nobody can rebuild, is how attention moved after each post.
LunarCrush processes up to 50M posts per hour across X, TikTok, Reddit, YouTube, and 10K+ news sources, and joins every post, creator, and topic to a forward attention curve: how engagement compounded over the next hours and days, how the topic spread across networks, and how it ranked against everything else competing for attention in that same window. Here is why that can't be scraped: a platform only ever serves the current cumulative count. To hold a post's hour-by-hour trajectory you had to be measuring while it happened. We were, at firehose scale, for seven years. Price is the final leg of the join and we treat it honestly: price history is public, so it's context we attach, never the moat we claim. Add years of AltRank and Galaxy Score history and you get something rare in machine learning: a massive dataset where every example carries a ground-truth answer to the question institutions pay us for: "did this social activity matter?"
Up to 50M posts/hour, multi-network, with creator identity and topic graph attached. Multimodal from the source: TikTok and YouTube are audio and video our text-first pipeline currently underuses. Inkling processes audio natively.
Sentiment, spam/bot classification, topic assignment, creator influence: labels our pipeline has been generating continuously. Every one is a training example we already paid for.
Post → next-24h/7d attention trajectory. Topic → cross-network spread and rank against everything else live that hour. Creator call → realized attention accuracy. Supervised targets that exist only because we were polling while it happened; no archive of them exists to buy.
AltRank and Galaxy Score encode a decade of institutional feedback about what signal looks like. Used as training targets, they distill that judgment into weights, a second life for our core IP.
One guardrail, stated up front because it shapes everything downstream: we train on our derived labels, metrics, and outcome joins, not on redistributing raw platform content. Selling scores from a model trained on our own labels is the same legal shape as selling AltRank today. This keeps us clean with X, TikTok, and Reddit terms while preserving all the value.
Within hours of v1, the sharpest objection this plan will face surfaced internally: everyone has posts, everyone has price history, so a post-to-price training set is something any lab or fund can rebuild. Correct. Here is what that kills, what it doesn't touch, and what it sharpens.
What it kills. Any version of this plan whose product is "a model that predicts price from posts." A scraped post corpus joined to free candles has been public alpha-mining territory for a decade; every serious quant desk runs an in-house version, and on that dataset dimension we hold no advantage over our own customers. Nothing in this paper rests on price-labeled training data being scarce, and v1 was wrong wherever it implied otherwise.
What it doesn't touch. The attention curve. Price prints continuously on an exchange whether or not anyone records it; anyone can fetch 2021's candles today. Engagement doesn't work that way. A platform serves only the current cumulative count, so a post's hour-by-hour attention trajectory exists only if someone was polling it while it happened, and it lives nowhere except panels like ours. The labs' trillions of post tokens carry, at best, one frozen engagement number from crawl day. The curves cannot be bought back at any price, because they were never stored anywhere else.
What it sharpens. The commercial posture. Funds neither trust vendor alpha nor need it; they build their own price models and buy the inputs they can't. So we sell the leg they cannot rebuild: calibrated attention forecasts, attention-derived scores, crowding and attention-risk signals their models consume. And the probe now carries a control arm built to falsify our own premise: the same base model fine-tuned once on public-reconstructible labels (posts joined to price moves) and once on ours. If our labels don't beat public ones on held-out evals, the moat is real but commercially worthless, and we stop.
This is not a moonshot proposal. Each phase is small, has a measurable gate, and only unlocks the next phase by winning on evals. Failure at any gate stops spend.
Export a few thousand outcome-labeled examples (≈1 day of Dan's time; the agent fleet does the rest). Fine-tune Inkling-Small on Tinker during the discount window. Benchmark ruthlessly on three tasks: classification cost-per-correct-label vs. our current provider rotation, calibrated next-24h attention forecasting vs. a well-prompted frontier model, and signal quality vs. the AltRank baseline on held-out history. Task 2 runs with a control arm: the same base fine-tuned on public-reconstructible labels only (post text joined to price moves), so the eval directly tests whether our proprietary attention labels add anything a competitor couldn't rebuild.
Gate: beat frontier on cost, match on quality, AND beat the public-labels control, or stopDeploy the winner on serverless inference for pipeline classification. Immediate payoff is margin and reach: sentiment and bot-detection across the full firehose at a fraction of rented-API cost, including networks we currently sample thinly.
Gate: production stability + measured cost win for one monthProductize as a "score anything" API for institutional customers: feed it any text, post, filing, or feed and get calibrated LunarCrush-framework scores and attention forecasts. Deliberately attention-native, not price prediction: the input leg funds can't rebuild, attached to data they already buy. Legal review of the ToS posture ships before this does.
Gate: 2 paying design partners from the existing pipelineThe Rebuy playbook applied to weights: time-boxed exclusive access to a specialized signal model for one fund. We already know this deal shape, this buyer psychology, and this contract structure.
Gate: Phase 2 evals + a fund at the tableRe-ranked after the objection: defensibility times nearness of cash. The surest money first; the biggest checks once the probe proves the labels. Enterprise at $95K/yr is the reference deal. None of these is "we sell price predictions."
Replacing rented frontier-API classification with our own 12B-active model cuts unit cost and makes full-firehose coverage affordable. Immune to the commoditization objection entirely: this pays even if nobody ever buys a score, and probe task 1 measures it directly.
Attention-native endpoints ("score anything" in our framework, calibrated attention forecasts) sold to buyers who already pay for our data. We supply the input their in-house price models can't rebuild. Same buyers, same sales motion, higher ACV; five institutional attaches is a meaningful step toward the 2x-to-$2M goal.
A specialized attention-forecast model licensed exclusively to one fund for 12–18 months. Funds pay multiples for inputs nobody else holds, and this one cannot be reconstructed from public data at any budget. CommercePulse economics applied to weights; gated on the probe's control arm proving the labels carry signal.
A new top tier where the product isn't data access but model access: custom scores, custom universes, fine-tuned per customer on our base. Positions us above the $95K tier rather than discounting it.
"LunarCrush has a proprietary model trained on seven years of attention-outcome data" changes how acquirers and investors compute our multiple. Data companies get data-company multiples; model companies do better.
The deeper reason to do this isn't any single revenue line. It's that a house model converts our operating exhaust into layered, defensible IP. Eight layers, from bedrock to flywheel. Tap to expand.
A white paper you can trust names its own failure modes. Here are ours, with the mitigation and the kill rule attached.
The entire case rests on evals. If a well-prompted GPT-5.5 with our data dictionary matches our fine-tune on quality at acceptable cost, the house model is vanity. Rule: two failed probe rounds → stop, revisit in 6 months. Bases keep improving; the corpus compounds either way. A failed probe costs ~$2K and teaches us the gap.
The sharpest internal objection, promoted to a formal gate. If a model fine-tuned on public-reconstructible labels (posts joined to price history) matches one fine-tuned on our attention curves, the moat is real but commercially worthless. The probe's control arm tests this head-to-head; losing it is a kill, not a caveat, no matter how well the other tasks score.
We never train for content redistribution: derived labels and outcome joins only, selling scores not posts. Same legal shape as AltRank. Formal legal review is a hard gate before the Phase-2 enterprise endpoint ships.
The probe is designed to need ~1 day of Dan (data export). Fine-tuning is managed (Tinker), evals run on the agent fleet, and nothing here touches the Rebuy or institutional pipelines, which stay priority one and two.
Social language drifts in weeks. Quarterly refreshes from the always-growing corpus are the plan, and the same drift is a barrier for anyone trying to copy a snapshot of us.
Inkling might be leapfrogged next quarter. Fine: the corpus, evals, recipes, and benchmark all transfer to any future base. We're betting on our data, not on Thinking Machines.
If the probe wins, we have a new product line with five revenue paths and an IP stack that compounds. If it loses, we spent lunch money to learn exactly where the frontier still beats us, and our dataset kept growing the whole time.