Experimental · NVIDIA Nemotron

A language model in the trading seat

Skydweller is rules-based: it fires when five gates pass. This is the other way of running AlgoTrada — an NVIDIA Nemotron-3-Nano-4B, fine-tuned on our own corpus and running on our own GPU, choosing the direction while deterministic code keeps every price, size and veto.

It has been placing real orders on a live book since 2026-06-1570 days, 23,582 fills, around 337 a day. This page is the experiment's open notebook, including the parts that do not flatter it.

NEMOTRON4BCUBICLESELF-HOSTED
Live since
2026-06-15
70 days
Fills
23,582
100% taker
Win rate
54.31%
8,607W / 7,241L
Realized P&L
$2,588
the only banked number
Read this before the rest

Start with what it was given: nothing. The venue enforces no collateral, no margin and no balance, so there is no funded account here and no starting capital to measure a return against. Any percentage figure would be invented. That is why this page reports absolute dollars, fills and a win rate — and never a return.

What it is carrying is real, though. It trades in $25 clips but has accumulated $157,316 of cost basis across four open positions, because it opens far more often than it closes. On a funded venue that position would have to be collateralised; here it simply grows.

Marked to market the agent is up roughly $47,288. About 95% of that is unrealized — open long positions in a market that rose. That is exposure, not edge, and it reverses as fast as it accrued.

The defensible number is the one in the table above: 54.31% across 23,582 fills, paying the spread on every single one. We publish this as an experiment. We are not selling its returns.

How the seat actually works

The model is not left alone with the money. It answers one question, and arithmetic does the rest.

01 · Nemotron-3-Nano-4B

The model decides

Nemotron sees the current structure — where price sits against the session range, whether the correlated book agrees, how volume compares to average — and answers with one word: BUY, SELL or HOLD.

02 · deterministic

Arithmetic prices it

Every level the model suggests is discarded. Entry, stop and target are recomputed from the live book and the session's own geometry. Position size is fixed. A structural risk gate can veto the trade outright.

03 · public receipt

The venue records it

The order goes to the Taifoon CLOB under a public attribution. Fills, positions and P&L are the venue's record, not ours — anyone can read them without our permission.

The organising idea behind the structure it reads is Goldbach confluence — the same family of session geometry Skydweller uses to decide where a level matters. The model was fine-tuned until it reasons in that language natively, which is why its notes read like a desk analyst rather than a chatbot.

What it actually says, unedited

Every order carries the model's own note. These are verbatim from the decision log — including the last one, where the note contradicts the action it took.

TimeMarketCallThe model's note
09:39:15BTCBUYsetup: BULLISH SMT, zone: gap low, action: BUY
09:40:17SOLBUYsetup: gap high, close near range low, volume 1.0x avg, signal: bullish
05:23:38NQSELLsetup: gap low, close near eq, volume 0.0x avg
09:39:35ETHSELLclose near range high, actual outcome: HOLD

The note is a read, not a proof. It is generated in the same pass as the decision and is not checked against it, which is why it can disagree with the trade it accompanies. We show those cases rather than filtering them out.

What it is holding right now

Four open positions, three of them long. This is where almost all of the headline number lives — and why it is not a result yet.

MarketSideSizeAvg entryMarkCost basisUnrealized
SOLlong932.1876.97101.74$71,751+$23,086
BTClong0.66365,972.7980,650.1$43,740+$9,731
ETHlong19.7671,904.332,503.44$37,642+$11,842
NQshort0.14329,264.2829,228.25$4,183+$5
Total carried$157,316+$44,700

Every long was opened well below where the market now trades, and none of it is banked. Against that, $2,588 has actually been realised in 70 days. The gap between those two numbers is the entire argument of this page.

It trades constantly

This is not an agent that takes a position a week. It works four books around the clock in small, uniform clips — roughly 337 fills a day and $7,631 of turnover a day, every one of them crossing the spread to get filled.

Turnover
$534,203
over 70 days
Fills / day
337
across four books
Avg clip
$22.65
deliberately small
Maker fills
0
100% taker, always pays

Where the volume goes

Turnover is spread almost evenly across the four books. The direction mix is not — three of them it reads long, one it reads both ways.

BTC$143,871 · 5,759 fills · 82.3/day · avg $24.98
33B / 12S

Largest book by turnover. Reads the floor far more often than the ceiling.

NQ$141,837 · 5,673 fills · 81/day · avg $25.00
21B / 25S

The only two-sided book — the one market it shorts with conviction.

ETH$130,860 · 5,260 fills · 75.1/day · avg $24.88
33B / 10S

Anchors on equilibrium; the most patient of the four.

SOL$117,636 · 6,890 fills · 98.4/day · avg $17.07
29B / 8S

Most fills, smallest clips — the book where orders get partially filled.

The odd one out is SOL: the most fills but the smallest average clip ($17.07 against a $25 target), which means its orders are landing partially filled. The first bar is share of turnover; the second is the buy/sell split.

What is wrong with it

An experiment that only publishes its wins is advertising. These are the three findings that most change how the record above should be read.

It can only sell at low conviction

Every buy it makes is graded top-tier; almost every sell is graded a rung below, and its stated confidence splits the same way. It does sell — when price sits high it is close to even-handed — but the grade and the direction are entangled, so a short can never carry the conviction a long does.

Its confidence means nothing

511 of 589 decisions came in at 94–95%, including ones its own note marks as the wrong call. A number that never varies carries no information, so we never show it without that context.

Roughly 9% of orders never land

Venue timeouts. A signal feed that quietly dropped those would overstate itself, so failed orders are published alongside filled ones and counted against the fill rate.

The first of those is why the record is not yet a claim. We now log the market structure behind every decision, and it shows the long lean is largely a matter of where price sat: shown a market trading high, the agent splits its calls almost evenly; shown one trading low, it buys nearly three times out of four. Until it can hold a short with the same conviction it holds a long, these results measure a rising market as much as they measure the agent — and we will not present them as anything else.

Nothing here asks you to trust us

The agent trades under a public attribution on a public venue. Every number on this page — the fills, the turnover, the win rate, the open positions — is the venue's own record, readable without our permission and without our cooperation.

One caveat worth knowing: the venue marks this actor an “LLM” by matching the name its orders submit, and a name is something anyone can claim. The label is not the proof. The trade history is.

What happens next

In order, because each step depends on the one before it. The first is not a feature — it is the correction that has to land before any of this record means anything.

01

Instrument every decision

in progress

Today the model's reasoning is only visible when it happens to mention it in prose. The structure it was shown, the grade it assigned, and its confidence need to be recorded on every single decision, whether it talks about them or not. Without that, none of the defects above can be measured properly — only guessed at.

02

Teach it to short with conviction

the blocking fix

The training corpus is already balanced — 26,140 short-labelled setups against 23,688 long, graded by realised outcome across four horizons. The agent's long lean is therefore not inherited from the data; it comes from how grade and direction were coupled during fine-tuning. Retrain with that coupling broken, on the corpus we already hold.

03

Re-baseline on the fixed agent

the honest reset

The record on this page describes the current, one-sided configuration. Once the agent can express conviction in both directions, that record stops applying and the count starts again — a fresh window, both-sided, before a single performance claim is made.

04

Condition on confluence, per minute

datalake

We keep a minute-by-minute record of structural confluence across every book, sitting alongside the labelled outcomes and sharing the same timestamp. Joining the two lets the model be trained on what the structure actually was, not on a summary of it — the single largest unexploited asset behind this agent.

05

Publish the signals as they happen

built, not yet live

Entries and exits, each carrying the model's own note and its lifetime record, with failed orders published alongside filled ones so the fill rate can never be flattered. The publishing decision is deliberate and has not been taken yet.

06

Open the seat to any model

the platform

The venue can already drive an external endpoint through a replay of real tape and hand back a causality receipt proving every bar was answered. Point it at your own model instead of ours, and the same rails that produced this page produce yours. This is the version that matters — it needs no edge claim from us at all.

The model, as NVIDIA ships it

open weights

Everything below is from NVIDIA's own published model card. We run the released weights — we did not build the model, only the adapter and the machinery around it.

ModelNVIDIA-Nemotron-3-Nano-4B
ArchitectureDense hybrid Mamba-2 + MLP, four attention layers
LineageCompressed from Nemotron-Nano-9B-v2
Context262,144 tokens
ReasoningToggleable; we run it off for latency
Where it runsOne GPU we own. No external API.

A 262K context is far more than a single trading decision needs. The headroom is what makes step 04 above possible — a whole session of structure can be put in front of the model at once, rather than a summary of it.

How a 4B model is made reliable enough to trade

A four-billion-parameter model is small. That is the point — it runs on hardware we own, answers in seconds, and costs nothing per decision. The trade-off is that it cannot be trusted with arithmetic or with its own certainty, so the engineering around it is built to take neither.

  • Ask for one thing.The model answers a single question — direction — in a fixed, machine-checkable shape. Every number it volunteers is discarded and recomputed.
  • Adapt, don't retrain. The base model is left alone; the domain lives in a small trained adapter on top, which can be rebuilt in hours and swapped without touching the rest.
  • Speak one dialect. Fine-tuning on our own graded corpus is what makes it reason in the language of session structure rather than in generic market talk.
  • Never trust the confidence.Small models are badly calibrated, and this one is measurably so. Conviction has to be earned from structure, not accepted from the model's own claim about itself.

Why publish an experiment that has not proven itself

Because the interesting result was never the P&L. It is that a 4B model, small enough to run on one of our own machines, learned to read session structure in our own language and has held a live seat for 70 days without supervision.

Whether it can trade is still open. What is already settled is that the record is checkable — and a record you can check is worth more than a return you have to believe.