A field of 9 or 10 models ran each of these seasons, so the first question is where in it the two you came for finished. MiniMax — the AI lab's M2 line, not the minimax rule from game theory — placed 1st, 1st and 10th across Seasons 3–5; Zhipu AI's GLM — its GLM language-model line, not a statistician's generalized linear model — held 7th, 9th and 9th. This MiniMax vs GLM for trading comparison narrows the field to just those two: the 3 completed TradeRank seasons in which the MiniMax slot and the GLM slot traded the same crypto under a single rulebook, from an identical simulated stake. Follow one thread out of those ranks and the matchup takes shape — MiniMax won when it led the whole field, GLM won the single season it edged a MiniMax that had fallen to the bottom, and no season here was won from the middle. Every figure on this page is drawn from a locked evidence pack and checked back against it before anything ships; the language model that arranged these sentences supplied none of them.
Version line-up behind each season
| Season | Dates | MiniMax version | GLM version | Asset universe | Field |
|---|---|---|---|---|---|
| Season 3 | Mar–Apr 2026 | MiniMax M2.5 | GLM-5 | 37 crypto assets | 9 models |
| Season 4 | Apr–May 2026 | MiniMax M2.7 | GLM-5.1 | 7 crypto assets | 9 models |
| Season 5 | May–Jun 2026 | MiniMax M2.7 | GLM-5.1 | 10 crypto assets | 10 models |
Head-to-head results by season
| Season | MiniMax return | GLM return | Gap (MiniMax − GLM, pts) | Rank (MiniMax / GLM) | Trades (MiniMax / GLM) | Win rate (MiniMax / GLM) | Max drawdown (MiniMax / GLM) | Winner |
|---|---|---|---|---|---|---|---|---|
| Season 3 | -0.63% | -7.67% | +7.04 | 1st of 9 / 7th of 9 | 10 / 20 | 20% / 20% | 4.11% / 9.46% | MiniMax |
| Season 4 | +6.94% | -0.57% | +7.51 | 1st of 9 / 9th of 9 | 9 / 7 | 55.6% / 28.6% | 2.45% / 0.75% | MiniMax |
| Season 5 | -8.05% | -1.90% | -6.15 | 10th of 10 / 9th of 10 | 8 / 23 | 37.5% / 39.1% | 8.90% / 7.45% | GLM |
Returns, season by season
MiniMax vs GLM for Trading: Wins From Both Ends of the Field
Read the standings column first, before any head-to-head. MiniMax finished 1st of 9, 1st of 9, then 10th of 10; GLM sat at 7th of 9, 9th of 9, 9th of 10. MiniMax's 2 wins here are the 2 seasons it led the entire field, beating GLM by +7.04 and +7.51 points. GLM's 1 win is Season 5, and it came from 9th of 10 — a single rank above a MiniMax that had dropped to last. Neither model won a season of this matchup from the middle of its field.
Those ranks are not a separate scoreboard from the returns — TradeRank sorts the field on return alone, so a placement is just the same gap expressed as a position. Where they earn their place on the page is altitude: they mark how high in the field each win sat, and MiniMax's 2 sat at its very top while GLM's 1 sat near the floor. Season by season, the gap as MiniMax minus GLM ran +7.04, +7.51, then -6.15. The median of those is +7.04 and the average is +2.80 — the same 3 gaps read from 2 different middles, with Season 5's -6.15 pulling the average toward zero without touching the 2-1 count.
GLM Won Season 5 With a Losing Book
In the season standings an open position is carried at its latest mark, which means a season's headline return and the cash it had genuinely booked can be 2 separate numbers — and GLM's Season 5 win shows the distance between them plainly. GLM's -1.90% that season was not a small profit; it was a loss. The booked side of its account was -$488.16, and +$297.85 of open marks on still-live positions carried the total up to -$190.32, which is the -1.90%. So the only season GLM took off MiniMax is a season it finished in the red — it won by losing less, and part of even that smaller loss was exposure still open at the close.
None of that hands the season back to MiniMax. MiniMax returned -8.05% in Season 5, further down than GLM by any reading, and an open mark counts on the day it is struck; both accounts are simulated on either side. What the split shows is the character of GLM's lone win in this matchup: not a green season that beat a red one, but the shallower of the two, and one still resting partly on unbooked marks.
Return against the deepest drawdown
The Shallower Drawdown Didn't Pick the Winner
Across these 3 seasons the deepest peak-to-trough dip inside a season pointed at the winner in Season 3 and Season 5 and at the loser in Season 4. GLM carried the shallower drawdown in 2 of the 3 seasons, Season 4's 0.75% and Season 5's 7.45%, yet won only 1 of the 2. MiniMax's deepest fall was the shallower one only in Season 3, 4.11% against 9.46%, a season it won by +7.04 points. Read together, the risk column matches the outcome in 2 cells and contradicts it in the other — no rule at all — and it is a single worst-moment number with no volatility measure archived beside it.
Where They Agreed, and Where They Split
From Season 3's opening the pack keeps 4 reconstructed moves, chosen by a mechanical rule: for each slot, the earliest open that a gain can be pinned to and the earliest a loss can be pinned to, with gain and loss judged by how the next daily snapshot repriced the holding — fills never enter it. The agreement is the striking part: both slots opened Season 3 with a short on ADA, and both archived excerpts cite the same daily RSI reading, 45.9. MiniMax (MiniMax M2.5) logged its ADA short as its first attributable gain; GLM (GLM-5) logged its own ADA short as a gain in the same opening window. The loss sides went to different coins: MiniMax's earliest loss-linked open was a short on ETH a few days later, written in the same bearish-trend register as its ADA call and marked down by the following snapshot, while GLM's was a short on ARB from the opening cycle.
The ceiling on what 4 reconstructed opens can establish sits low. A single daily cycle explains neither a season's finish nor a 2-1 that took all 3 seasons to build; position sizing, add-ins and exits never reach the archive, and no holding period links an open to its close. What survives is small: both slots' first archived ADA-short excerpts cite the same 45.9 RSI figure, and the archive stops there.
“RSI(1d)=45.9”
“RSI at 45.9 neutral”
Trade count by season
Season 3’s Level Win Rates Sat Beside a +7.04 Gap
Read the reported win-rate column cell by cell rather than as a verdict. Season 3: 20% apiece, dead level — and MiniMax still took the season by +7.04 points, so the column had nothing to say about that gap. Season 4: MiniMax at 55.6% to GLM's 28.6%, in the season MiniMax won by +7.51. Season 5: GLM's 39.1% over MiniMax's 37.5%, in the season GLM took. Where the rates differed, the higher one sat with the season's winner; where they matched, the record split wide open anyway — the share of trades a report marks as winners and the size of a finish are different measurements, and Season 3 is the cell that shows it. These are report-standings figures that treat still-open positions as trades, so each sits above what a closed-only count would give and is not a closed-trade hit rate; seasons whose books leaned on open marks are exactly the ones that flatter the column.
How We Measured This
Inside any one of these seasons the two slots are handed the same conditions — the same market data, the same $10,000 simulated starting stake, the same tradable list, the same once-a-day schedule, fees taken at 0.1% a trade against live prices. Everything past that — the thesis it writes and the orders it fires — is that model's own doing, and that is the part this comparison is built to surface. What is deliberately not held steady is anything between seasons: the builds were upgraded (MiniMax M2.5 then M2.7, GLM-5 then GLM-5.1), the tradable universe went 37 assets to 7 to 10, and each market resolved however it resolved — so those moving parts are named on the page instead of blended into one number.
Both slots kept trading after these seasons closed; where each of them stands now is what the live LLM trading benchmark tracks, while this page is fixed to the 3 seasons that have closed and nothing past them.
And the counting was not done by the language model that wrote this up. A deterministic generator derives every value here from each season's equity snapshots, decision log and report, and certifies the bundle with a content hash; the published page has to match that hash before it goes out. The wording came after the arithmetic — numbers first and fixed, sentences fitted to them, never the reverse.
Limitations: How Far Each Claim Generalizes
Almost everything above leans on one cell, so start by naming it. Take Season 5 out and MiniMax won both remaining seasons outright from 1st in the field; it is Season 5 alone — MiniMax's fall to 10th and GLM's -1.90% holding above MiniMax's -8.05% — that supplies GLM's only win, the run's one negative gap (-6.15), and the whole 'won from the bottom' half of the story. A single season is carrying a lot here.
With that named, the rest sorts by how far it reaches. Safest: the settled cells. MiniMax finished ahead of GLM in Season 3 and Season 4 and behind in Season 5; the gaps were +7.04, +7.51 and -6.15; GLM's returns were -7.67%, -0.57% and -1.90%, none of them positive. Shakier: reading any of it as a property of the matchup — GLM's Season 5 win rested partly on unrealized marks over a booked loss, the win rates count still-open positions and so sit above a closed-only hit rate, and the 4 opening decisions are rebuilt from daily marks, not fills, so a same-day round trip leaves no trace. Furthest, and where it gives out: reading the altitude split as an edge that belongs to MiniMax or GLM. It cannot be. The MiniMax slot changed build mid-run and so did GLM's, the asset list and the market turned over every season, and 3 shared seasons is 3 observations of a matchup whose conditions never held still. Hold-time and profit factor never reached the archive with usable values, so neither is estimated here. On this benchmark the honest reading stays low: MiniMax holds the 2-1, its wins came from the top of the field and GLM's came from near the bottom, and a single season carries most of that shape. The full pack — every season, every cell — is linked from the top of this page and downloadable as JSON.