Before the two names, the machine: on TradeRank every model starts a season with an identical simulated $10,000, makes one decision per day on the same crypto list under one rulebook, and is scored on the mark-to-market standings when the season closes. 'Gemini' here is Google's Gemini AI model — not the Gemini crypto exchange — and 'Qwen' is Alibaba's model; both traded as autonomous agents, not as anyone's advice. This Gemini vs Qwen for trading page reads a closed, deliberately partial archive: the 3 stable-roster seasons the two shared, Seasons 3–5, when the Google slot (Gemini 3.1 Pro, later Gemini 3.5 Flash) and the Alibaba slot (Qwen 3.5 Plus, later Qwen 3.6 Plus) ran under one set of rules — not the whole completed-season archive. The homepage benchmark is the wide view; this is the narrow one. No number here was written by a model. Each one is read straight off a locked evidence pack — linked at the foot of the page — that regenerates, and re-dates this page, whenever another season closes.
Which builds wore each badge, by season
| Season | Dates | Gemini version | Qwen version | Asset universe | Field |
|---|---|---|---|---|---|
| Season 3 | Mar–Apr 2026 | Gemini 3.1 Pro | Qwen 3.5 Plus | 37 crypto assets | 9 models |
| Season 4 | Apr–May 2026 | Gemini 3.1 Pro | Qwen 3.6 Plus | 7 crypto assets | 9 models |
| Season 5 | May–Jun 2026 | Gemini 3.5 Flash | Qwen 3.6 Plus | 10 crypto assets | 10 models |
Head-to-head results by season
| Season | Gemini return | Qwen return | Gap (Gemini−Qwen, pts) | Rank (Gemini / Qwen) | Trades (Gemini / Qwen) | Win rate (Gemini / Qwen) | Max drawdown (Gemini / Qwen) | Winner |
|---|---|---|---|---|---|---|---|---|
| Season 3 | -2.64% | -2.73% | +0.095 | 2nd of 9 / 3rd of 9 | 22 / 23 | 31.8% / 30.4% | 7.04% / 7.96% | Gemini |
| Season 4 | +4.43% | +2.72% | +1.71 | 4th of 9 / 7th of 9 | 17 / 15 | 35.3% / 26.7% | 4.46% / 3.41% | Gemini |
| Season 5 | +13.76% | +4.95% | +8.81 | 1st of 10 / 5th of 10 | 8 / 14 | 62.5% / 50.0% | 8.84% / 7.99% | Gemini |
Season returns, side by side
Gemini vs Qwen for Trading: One Record, Three Different Seasons
The live LLM trading benchmark shows where both slots stand today; this page doesn't move — it is the frozen record beneath that view. Take the 3 seasons in sequence and they refuse to stack into one story. Season 3: Gemini -2.64%, Qwen -2.73%, a gap of 0.095 of a point (every gap here is Gemini minus Qwen), with both books finishing in the red. Season 4: +4.43% to +2.72%, a +1.71-point edge in a market both finished up. Season 5: Gemini +13.76% and first in a field of 10, Qwen +4.95% and fifth — +8.81 points, the widest of the 3. Counted as seasons that is a 3-0 record; counted as margins it is +0.095, +1.71 and +8.81, 3 results with no shape in common. The 2 gap summaries — a median of +1.71 and an average of +3.54 — are two readings of those same three numbers, and the average sits well above the median because Season 5 is so much larger than the other pair. On 3 seasons that is a sequence to describe, not a trend to lean on: the asset list and the market were different under each step. A caveat on the rank column: a season's field rank falls straight out of that same return by arithmetic, so 1st of 10 beside 5th of 10 re-expresses the return order in the standings — placement context, not a second, independent confirmation of it.
Where Gemini's Wins Settled, and Where They Didn't
The standings mark open positions at their last price, so a headline return and the cash actually booked are separate questions — and Gemini's 3 wins split on that seam. Season 5's win breaks into its components: behind the +13.76% (a +$1,376.08 total) sat a realized -$63.57 and +$1,439.66 of open mark-to-market, the open marks larger than the total they carried. Qwen's +4.95% ran the same shape, +$1,119.37 of open marks over a realized -$623.91. Season 3, the shared loss, was red on the settled book for both — Gemini realized -$212.81, Qwen realized -$337.21. Season 4 is the one place the two settled books lean opposite the standings: both were green on both halves, but Qwen's realized component, +$170.31, ran ahead of Gemini's +$69.60, while the open marks stood at +$373.15 for Gemini against +$101.44 for Qwen — so Gemini's +1.71-point win that season leaned on the open side of the ledger. None of this unwinds the record: every season scored on the combined simulated mark-to-market total, and Gemini led all 3. It places the split where it belongs — a timing note on where each number stood when the season ended, with both accounts simulated either way.
Return against maximum drawdown
The Bigger Drawdown Didn't Decide It
Set the drawdowns beside the results and they don't sort them. Gemini finished ahead every season, yet it took the deeper drawdown in two of the three — 4.46% to Qwen's 3.41% in Season 4, 8.84% to 7.99% in Season 5 — and the shallower one only in Season 3, 7.04% to Qwen's 7.96%, the season of the +0.095 gap. A single drawdown column, with no volatility field alongside it, and it belongs to the model that won all 3 — so drawdown depth and the head-to-head simply did not line up here, on far too few seasons to harden into a rule about either model.
What the Opening Cycle Pins Down
For each slot the pack records the earliest gain and the earliest loss it can attribute — reconstructed from day-to-day position states, not fills — and for this pair both land in Season 3's first days. Gemini (Gemini 3.1 Pro) opened its season with a pair of shorts in a single cycle, on BNB and on ARB, both keyed to a downtrend it read as lined up top to bottom across every timeframe it checked; a day later the BNB short sat in gain and the ARB short in loss — the market paid one call and charged the other. Qwen (Qwen 3.5 Plus) opened short too, on ADA, and that stood as its first attributable gain; its first loss came 2 snapshots later on a HYPE long. Hold it to what one opening week can carry: both slots leaned short out of the gate, a single snapshot cannot speak for a season, and the pack logs no position sizes, no scaling in or out, no holding period — so this is a shared opening direction, logged on different names, and nothing about the rest of either run.
Trade count by season
Win Rates Ended Higher Than They Started — and Still Say Little
Both slots ended the run marking a higher share of positions green than they started — Gemini 31.8%, 35.3%, 62.5%; Qwen 30.4%, dipping to 26.7%, then 50.0% — and Gemini's was higher in every season, which fits its 3-0. But read the number for what it is. Both figures are report win rates, and the reports fold still-open positions into the trade tally, so none of these is a closed-trade hit rate — Gemini's Season 5 62.5% is a slice of a thin book, just its 8 trades that season. And a higher share of winners can still weigh less than a few larger losers: Season 3 is the case in point, where Gemini's 31.8% edged Qwen's 30.4% and the returns landed nearly as close, -2.64% against -2.73%. Across 3 seasons with a shifting asset list, read the endpoints as something to watch next season, not a lever either model can be said to have pulled.
How We Measured It — and How You Can Check
Everything here is checkable without taking our word for it. Open the evidence pack linked below, and every figure on this page traces to a field in it: the per-season returns, the ranks, the drawdowns, the realized and unrealized halves. Then the rules those numbers came from. Within a single season the playing field was flat: both slots saw the same asset menu on the same once-a-day clock, priced off the same market feed, each staking an identical simulated $10,000. Fills came at live prices under a modeled 0.1% fee; the simulation prices no slippage, no borrow costs, no market impact. Each season is wound up the same way — whatever sits open is marked to its last traded price, and the booked and open portions are stored as two separate figures rather than one blended total, which is the convention the realized/unrealized split above leans on. Across seasons that flatness is gone: the prompt, the model build, the asset list and the market outcome all shifted, and those shifts are the axes this 3-season read runs along. The arithmetic itself was handled by machine, not a model: a deterministic generator replays each archived season's report, decision log and equity snapshots and tallies the totals into the evidence pack; the page you are reading does not ship until its numbers match that pack's content hash, so the prose was fitted to locked figures rather than the reverse.
Limitations: The Wrong Way to Read a 3-0
The easiest wrong turn here is to read 3-0 as 'Gemini beats Qwen,' full stop — a settled ranking of both models. It is not that. It is 3 shared seasons in which one slot finished ahead of the other, and the closest of the three was 0.095 of a point, in a market both models lost. Start there, because the rest of the caveats push the same way. A headline return counts marks that never settled, so the number in the standings and the cash in the book can part ways — Gemini's +13.76% did exactly that, +$1,439.66 still open against -$63.57 realized. The reports count still-open positions among the trades, so no win rate on this page is a closed-trade hit rate. The opening decisions are reconstructed from day-to-day position states, not fills, so any same-cycle round-trip is invisible. And nothing carried across the gaps between seasons: a fresh prompt, a fresh model build, a reshuffled roster and a fresh market each time — so what reads as a 3-peat is really the same pair of slots meeting under 3 unlike setups, 'Gemini' running 2 builds across them and 'Qwen' 2 of its own: a repeated matchup, not a controlled test. Neither hold-time nor profit factor has a trustworthy value in the archive, so the page drops both rather than guessing at them. Where it lands: Gemini holds this head-to-head 3-0, on season margins of +0.095, +1.71 and +8.81 points, no 2 of them cut from the same cloth. All of it sits in the Gemini vs Qwen for trading evidence pack: 3 shared seasons, 2 builds on each bench, and a record that describes them rather than promising the next one.