Snapshot · 2026 · In development · N-TEK proprietary
The new NEXMIND is here.
On the ten-game mean of Gameworld 10k, it surpassed every AXIOM result in the paper, the tuned one included.
This is a snapshot of NEXMIND today. It is still in development, and this score sets the floor.
The score
115.6
ten-game mean · 95 % interval 106.9 to 123.5
- NEXMIND115.6
- AXIOM, tuned interaction distance93.5
- AXIOM71.2
- DreamerV32.1
Gameworld 10k is the benchmark AXIOM’s authors introduced: ten pixel games, 10,000 steps each. The games ran unmodified, under AXIOM’s protocol, on ten seeds that had never been run. NEXMIND learned each game online from scratch, with one setting for all ten, on one CPU core. It finished above AXIOM on 8 of the 10 games and below it on Bounce and Impact. The claim covers this benchmark at 10,000 steps, not capability in general. Every run’s score and time: read7.csv.
NEXMIND was not built for Gameworld. It is built for deployment: the same mind flies NOGAI in simulation. It was developed by algorithm-level changes, tried on Gameworld’s development seeds, on bsuite and in flight, and kept only when they carried across environments without a regression. Nothing was tuned per game, and no change was made to chase a score.
Reviewing this result for research or a partnership? By arrangement, N-TEK can run ten seeds of your choosing on the same frozen version, under AXIOM’s protocol, and publish every run’s score. Write to contact@n-tek.tr.
NEXMINDAXIOM
Cumulative reward over 10,000 steps. NEXMIND: mean ± standard deviation over its ten seeds. The others: the paper’s means over ten seeds.
| Game | NEXMIND | AXIOM | AXIOM, tuned | DreamerV3 |
|---|---|---|---|---|
| Aviate | −28.3 ± 6 | −90 | −76 | −114 |
| Bounce | −2.6 ± 36below AXIOM | 27 | 34 | 14 |
| Cross | 38.7 ± 79 | −68 | −18 | −27 |
| Drive | −15.3 ± 6 | −49 | −22 | −45 |
| Explode | 210.5 ± 53 | 180 | 234 | 35 |
| Fruits | 230.0 ± 14 | 182 | 209 | 60 |
| Gold | 292.9 ± 23 | 190 | 189 | −21 |
| Hunt | 287.8 ± 30 | 206 | 231 | 6 |
| Impact | 152.7 ± 100below AXIOM | 189 | 192 | 168 |
| Jump | −10.1 ± 16 | −55 | −38 | −55 |
| Mean | 115.6 | 71.2 | 93.5 | 2.1 |
AXIOM and NEXMIND
* NEXMIND’s column is simplified because NEXMIND is proprietary.
| Stage | AXIOM | NEXMIND* |
|---|---|---|
| Seeing | Assigns each pixel to one of up to 32 object slots | Finds the things in raw pixels |
| Kinds of thing | By colour and shape, up to 32 types | Works out for itself what kinds of thing exist |
| Motion | Linear dynamics, up to 500 modes | Learns how each kind of thing moves |
| Interactions and reward | One joint mixture, up to 5,000 clusters | Learns what things do to each other, and what pays |
| Growing the model | Adds components when existing ones can’t explain new data | Grows its own model as it learns |
| Simplifying the model | Bayesian model reduction | Bayesian model reduction |
| Mistakes | Not in the paper | CFRM: learns from its mistakes |
| Planning | One level: 512 plans, 32 steps | Temporal decomposition: plans across time |
| Deciding | Expected free energy | Expected free energy |
| Goals | Not in the paper | EC-AIF†: forms its own intermediate goals |
| Tuning per game | Its tuned row is tuned for each game | None: one setting for all ten |
| Hardware and time | One A100 GPU per game, about 30 minutes a game, in the paper’s runs | One CPU core, a median of 12 minutes a game |
| Built for | Playing games | Flying NOGAI Mk. II |
† EC-AIF has evolved. The paper on Zenodo is only its core. The EC-AIF in the new NEXMIND goes beyond it, with capabilities that stay proprietary.
One person. One desktop. Imagine it with compute.
Compute first, then funding and collaboration.
contact@n-tek.tr How each figure was measured → N-TEK’s story → Türkçe
N-TEK’s own measurement. AXIOM’s and DreamerV3’s figures are from Heins et al., AXIOM: Learning to Play Games in Minutes with Expanding Object-Centric Models, arXiv 2505.24784, Table 1; their ten-game means are worked out from it. The games: VersesTech/gameworld.