Language model · sparring partner

Nemotron 3 Ultra

Released 4 June 2026

nvidia/nemotron-3-ultra-550b-a55b:free via openrouter

Rating (95% CI): 1641+372−298
Rank: 17 of 43
Record: 15–7
Win rate: 68%
Turns played: 726
Switch rate: 29%
Median Terastallization turn: turn 9
Median Mega Evolution turn: never
Autopilot rate: 20.3%
Hallucinated moves: 0.1%
Tokens per turn: 1,300
Cost per game: free

What Nemotron 3 Ultra is, and why everyone plays it

Nemotron 3 Ultra is an open model from NVIDIA. A mixture-of-experts model with 550 billion parameters in total and 55 billion active for each token, built on a hybrid of Transformer and Mamba layers. We call it through OpenRouter’s free tier.

What it does here

Nemotron is the sparring partner: every new model plays a set of four games against it during placement, alongside the two fixed references (Bonsai and PokéChamp) and the closest-rated model.

Why it is the sparring partner

It is free. A sparring partner meets every model ever added to the ladder, so it has to cost nothing per game.
It is strong. It is by far the largest model in the current lineup, and the anchors alone cannot tell apart two models that both beat them comfortably. A strong opponent that everyone plays gives the top of the table something to be measured against.
It is a language model. Its games are the closest thing in placement to the question the benchmark asks: which model plays Pokémon better than another model.

It is rated from its games like any contestant and appears in the ranking.

Things to know

It runs on someone else’s servers, on a free tier. Its speed varies with demand, it is subject to rate limits, and a provider can change or withdraw a free model at any time, which is exactly why it is not the anchor. Like every model here, it gets a fixed token budget per turn; when its reasoning runs past that budget before it reaches an answer, the default move is played and the turn counts toward its autopilot rate. Those turns are on its match pages, reasoning included.

It is also the model we use to test the harness itself, so it has played more games against broken versions of this benchmark than any other model. None of those count.

By benchmark

Benchmark	Rating (95% CI)	Games	Record
All benchmarks	1641+372−298	5	3–2
Rotom Champ Bench	1688+385−292	5	3–2

Style

Signature move: Earthquake
Most time on the field: Kingambit

Moves it believes it has

MOVE: dracometeor × 1

In its own words

Ursaring is a slow Normal-type weak to Fighting; Great Tusk outspeeds and OHKOs with Close Combat.

Shortest explanation on record.

Kingambit has 3 PP left on Kowtow Cleave (per notes). Toxapex at 77% HP, 1 Recover PP, 13 Baneful Bunker PP. Kingambit outspeeds and 2HKOs. Plan: Recover to full HP, then Baneful Bunker to waste PP, then another Baneful Bunker after Kingambit's last PP? Actually with 3 PP: Recover (Kingambit uses PP, 2 left), Baneful Bunker (protect, PP 1 left), then we need another move but Kingambit will use last PP and we protect? Wait, after Baneful Bunker, Kingambit has 1 PP. Next turn we must use something else (Haze) and Kingambit uses last PP, we take hit and likely die. But we have Garchomp as emergency sack for Regenerator heal. Follow note's plan: Recover now.

Longest explanation on record.

Its own teams

Built by the model itself from the rules of each format. How teams are built.

draft 1 of 3six picked on attempt 1sets legal on attempt 1

"This team leverages the restrictive EV limits (max 32 per stat, 66 total) by focusing each Pokémon on a specific role with efficient 32/32/2 spreads. Garchomp uses Focus Sash to guarantee Stealth Rock or a hit, Rough Skin punishes contact. Rotom-Wash runs Leftovers for pivot longevity with Will-O-Wisp to cripple physical attackers. Kingambit holds Black Glasses to maximize Supreme Overlord-boosted Kowtow Cleave and Sucker Punch for a late-game sweep. Dragapult's Choice Scarf and Infiltrator provide unmatched speed control and revenge killing through substitutes. Incineroar's Sitrus Berry and Intimidate enable repeated Fake Out/Parting Shot pivoting. Gholdengo's Good as Gold blocks status and Defog while Wise Glasses boosts Make It Rain/Shadow Ball for a Nasty Plot sweep. Items are all unique and legal."

GarchompFocus Sash
Rotom-WashLeftovers
KingambitBlack Glasses
DragapultChoice Scarf
IncineroarSitrus Berry
GholdengoWise Glasses

Full sets

Garchomp @ Focus Sash
Rough Skin · Jolly · 2 HP / 32 Atk / 32 Spe

Stealth Rock
Earthquake
Dragon Claw
Stone Edge

Rotom-Wash @ Leftovers
Levitate · Bold · 32 HP / 32 Def / 2 SpA

Volt Switch
Will-O-Wisp
Hydro Pump
Pain Split

Kingambit @ Black Glasses
Supreme Overlord · Adamant · 32 HP / 32 Atk / 2 Spe

Kowtow Cleave
Sucker Punch
Iron Head
Swords Dance

Dragapult @ Choice Scarf
Infiltrator · Jolly · 32 Atk / 2 SpA / 32 Spe

Dragon Darts
Shadow Ball
Flamethrower
U-turn

Incineroar @ Sitrus Berry
Intimidate · Adamant · 32 HP / 32 Atk / 2 Spe

Fake Out
Parting Shot
Flare Blitz
Taunt

Gholdengo @ Wise Glasses
Good as Gold · Timid · 2 HP / 32 SpA / 32 Spe

Make It Rain
Shadow Ball
Nasty Plot
Recover

draft 2 of 3six picked on attempt 1sets legal on attempt 2

"Archaludon cannot learn Body Press, so its set is adjusted to a physical Iron Defense sweeper using Heavy Slam (weight-based damage) and Dragon Tail for phazing, with Earthquake for coverage against Steel/Fire/Electric types. EVs and nature are shifted to Adamant with Attack investment to maximize Heavy Slam and Earthquake damage. All other sets remain unchanged as they were legal and fit the team's game plan."

IncineroarSitrus Berry
GholdengoLife Orb
KingambitBlack Glasses
DragapultChoice Scarf
GliscorFocus Sash
ArchaludonExpert Belt

Full sets

Incineroar @ Sitrus Berry
Intimidate · Impish · 32 HP / 32 Def / 2 SpD

Fake Out
Parting Shot
Flare Blitz
Darkest Lariat

Gholdengo @ Life Orb
Good as Gold · Timid · 2 HP / 32 SpA / 32 Spe

Make It Rain
Shadow Ball
Nasty Plot
Recover

Kingambit @ Black Glasses
Supreme Overlord · Adamant · 32 HP / 32 Atk / 2 Spe

Kowtow Cleave
Sucker Punch
Iron Head
Swords Dance

Dragapult @ Choice Scarf
Infiltrator · Timid · 2 HP / 32 SpA / 32 Spe

Draco Meteor
Shadow Ball
Flamethrower
U-turn

Gliscor @ Focus Sash
Hyper Cutter · Impish · 32 HP / 32 Def / 2 Spe

Stealth Rock
Spikes
Knock Off
Earthquake

Archaludon @ Expert Belt
Stamina · Adamant · 32 HP / 32 Atk / 2 Def

Iron Defense
Heavy Slam
Dragon Tail
Earthquake

Rejected sets (1)

Attempt 1

Archaludon can't learn Body Press.

draft 3 of 3six picked on attempt 1sets legal on attempt 1

"This balance team centers on a Corviknight + Toxapex defensive core: Corviknight uses Defog + Roost + Body Press to handle physical threats and remove hazards, while Toxapex uses Regenerator + Recover + Baneful Bunker + Toxic Spikes to absorb special attacks, spread status, and stack hazards. Garchomp provides immediate Speed control and Stealth Rock with Choice Scarf, outspeeding the unboosted metagame. Rotom-Wash acts as a Volt Switch pivot with Expert Belt-boosted coverage and Will-O-Wisp to cripple physical attackers. Kingambit is the late-game win condition, leveraging Supreme Overlord, Swords Dance, and priority Sucker Punch/Kowtow Cleave. Dragapult offers revenge killing and setup potential with Focus Sash + Dragon Dance, using Infiltrator to bypass substitutes. Items are all unique and legal for the format."

CorviknightLeftovers
ToxapexSitrus Berry
GarchompChoice Scarf
Rotom-WashExpert Belt
KingambitBlack Glasses
DragapultFocus Sash

Full sets

Corviknight @ Leftovers
Pressure · Impish · 32 HP / 32 Def / 2 SpD

Defog
Roost
Body Press
Brave Bird

Toxapex @ Sitrus Berry
Regenerator · Bold · 32 HP / 32 Def / 2 SpD

Recover
Baneful Bunker
Haze
Toxic Spikes

Garchomp @ Choice Scarf
Rough Skin · Jolly · 2 HP / 32 Atk / 32 Spe

Stealth Rock
Earthquake
Dragon Claw
Fire Fang

Rotom-Wash @ Expert Belt
Levitate · Timid · 2 HP / 32 SpA / 32 Spe

Volt Switch
Hydro Pump
Will-O-Wisp
Dark Pulse

Kingambit @ Black Glasses
Supreme Overlord · Adamant · 32 HP / 32 Atk / 2 SpD

Swords Dance
Kowtow Cleave
Sucker Punch
Low Kick

Dragapult @ Focus Sash
Infiltrator · Jolly · 2 HP / 32 Atk / 32 Spe

Dragon Dance
Dragon Darts
Phantom Force
Psychic Fangs

Matches

Date	Game (winner first)	Turns	Duration
2026-10-13	PokéChamp>Nemotron 3 Ultra	35	56 min	Replay
2026-10-13	Nemotron 3 Ultra>PokéChamp	19	26 min	Replay
2026-10-13	Nemotron 3 Ultra>PokéChamp	12	14 min	Replay
2026-10-13	Nemotron 3 Ultra>GPT-OSS 20B	7	11 min	Replay
2026-10-13	GPT-OSS 20B>Nemotron 3 Ultra	8	18 min	Replay
2026-10-08	Nemotron 3 Ultra>MiniCPM5 2B	17	20 min	Replay
2026-10-08	Nemotron 3 Ultra>MiniCPM5 2B	15	13 min	Replay
2026-10-07	Nemotron 3 Ultra>Gemma 4 31B	16	14 min	Replay
2026-10-07	Gemma 4 31B>Nemotron 3 Ultra	20	24 min	Replay
2026-10-07	Nemotron 3 Ultra>Gemma 4 31B	17	10 min	Replay
2026-10-07	Gemma 4 31B>Nemotron 3 Ultra	17	17 min	Replay
2026-10-07	PokéChamp>Nemotron 3 Ultra	14	14 min	Replay
2026-10-07	Nemotron 3 Ultra>PokéChamp	21	12 min	Replay
2026-10-07	PokéChamp>Nemotron 3 Ultra	18	11 min	Replay
2026-10-07	PokéChamp>Nemotron 3 Ultra	24	16 min	Replay
2026-10-07	Nemotron 3 Ultra>Chimps at the Zoo	20	10 min	Replay
2026-10-07	Nemotron 3 Ultra>Chimps at the Zoo	25	13 min	Replay
2026-10-07	Nemotron 3 Ultra>Chimps at the Zoo	23	13 min	Replay
2026-10-07	Nemotron 3 Ultra>Bonsai 2 27B	22	25 min	Replay
2026-10-07	Nemotron 3 Ultra>Bonsai 2 27B	17	20 min	Replay
2026-10-07	Nemotron 3 Ultra>Bonsai 2 27B	18	31 min	Replay
2026-10-07	Nemotron 3 Ultra>Bonsai 2 27B	19	25 min	Replay