Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarNeurohive→ original

MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms

A world model is a neural network that draws each frame from your key presses, with no game engine behind it. Neurohive tested seven such demos in the browser, including MIRA running Rocket League. MIRA's authors claim 20 frames per second, but ping (server response time) in the test ranged from 190 to 460 ms. So judge these models by control latency, not by frame rate.

AI-processed from Neurohive; edited by Hamidun News
MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
Source: Neurohive. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceNeurohive ↗
Editor-in-Chief
Key metric #1MIRA and LingBot-World 2.0 do not respond to WASD on a Russian keyboard layout
Key metric #2In LingBot-World 2.0, the lantern lit up 1.5-2 seconds after the click
Agora-24 players and 16 bots, the shot fired right after the click, no account needed

Imagine a game with no game engine: no level map, no physics, no pre-drawn models. You press a key, and a neural network draws the next frame on its own. Such systems are called world models: given the current scene and your actions, they predict what happens next.

The Neurohive editorial team tested seven browser demos. The most telling result came from MIRA, a neural-network version of Rocket League. Its authors claim 20 frames per second, which is the rate at which the model renders the picture.

But in testing, ping, meaning the time between your key press and the server's response, ranged from 190 to 460 milliseconds. That is almost half a second, enough time for the car to drive away from the ball. For both players and developers, what matters is control latency, not an impressive frame-rate figure.

What the best demos got right

Agora-2 responded best. It recreates a dungeon run from Diablo II. A single world holds up to 20 participants; in the preview, that is four humans and 16 bots. The model remembers where the heroes and monsters stand and shows the cave through each player's eyes: a teammate at the entrance may not yet see the enemy you are already fighting. The test author walked away from the group, came back, and recognized a bend in the cave by its walls and the map. The shot fired right after the mouse click. The demo is free and requires no account, but the fights quickly become repetitive.

Odyssey-3 works differently: you describe a place in text, and the neural network renders a walk through it. For a prompt about a black cat with a mechanical paw in a neon cyberpunk city at night, it produced a street close to the description. Neon reflected in the wet asphalt, and turns responded immediately. However, the cat's fur shifted slightly, and the free session is limited to five minutes.

Where it all breaks down

MIRA was trained on roughly 10,000 hours of bot-versus-bot matches. The teams at General Intuition and Kyutai built it with the participation of Epic Games. The match is two-on-two and is shown from four angles at once, so the same goal is visible from different ends of the field. But when the test author drove the car herself, she found it hard to keep track of the ball, and because of the lag she could not reach it in time.

In LingBot-World 2.0, it took one and a half to two seconds from highlighting a button to the lantern lighting up, and about a second to the blue flash of the shield. The world built from a reference image of a witch in a forest looked very close to the original, but the picture jerked and froze while walking. The ready-made preset lagged just as badly. The authors themselves admit that the model only runs at full strength on their own hardware.

Oasis, trained on Minecraft recordings by Decart and Etched, remembers its surroundings poorly. Turn away from a dirt wall and look back, and a wasteland may be in its place. Up close, the blocks blur, and it is unclear whether you managed to break even one.

In PixVerse R2, an interactive novel in the style of Van Gogh looks sharp, with individual brushstrokes visible. But the 320 starting points were enough for only one five-minute generation, and the text sometimes appeared before the picture changed.

A small thing that breaks the demo

MIRA and LingBot-World 2.0 recognize keys by character, not by position on the keyboard. On a Russian layout, instead of W they receive the Cyrillic letter "ts" (the letter on the same key as W) and do nothing. The car stands still, and the character does not walk forward. The spacebar and arrow keys do not depend on the layout, so in MIRA at first the only things that worked were jumping and driving with the arrows. Switch to an English layout before launching. If you are building your own browser demo, bind controls to key position, not to the character.

Why this matters

Generated worlds remain impressive experiments rather than games. The best-known world model, Genie from Google DeepMind, is not in the roundup: Project Genie, built on Genie 3, is available only with a Google AI Ultra subscription, and Russia is not on the list of supported countries. If you are evaluating such models for a product, check three things: the real latency from key press to frame, not the frame rate from a press release; how well details hold up over a long session; and the terms of free access, such as the timer, account requirement, and points limit.

For now, Agora-2 leads in responsiveness and recognizable imagery, while for MIRA the best picture does not make up for a ping of hundreds of milliseconds.

Story

Key Questions Answered

What the best demos got right?

Agora-2 responded best. It recreates a dungeon run from Diablo II. A single world holds up to 20 participants; in the preview, that is four humans and 16 bots. The model remembers where the heroes and monsters stand and shows the cave through each player's eyes: a teammate at the entrance may not yet see the enemy you are already fighting. The test author walked away from the group, came back, and recognized a bend in the cave by its walls and the map. The shot fired right after the mouse click.

Where it all breaks down?

MIRA was trained on roughly 10,000 hours of bot-versus-bot matches. The teams at General Intuition and Kyutai built it with the participation of Epic Games. The match is two-on-two and is shown from four angles at once, so the same goal is visible from different ends of the field. But when the test author drove the car herself, she found it hard to keep track of the ball, and because of the lag she could not reach it in time.

What are the key details on A small thing that breaks the demo?

MIRA and LingBot-World 2.0 recognize keys by character, not by position on the keyboard. On a Russian layout, instead of W they receive the Cyrillic letter "ts" (the letter on the same key as W) and do nothing. The car stands still, and the character does not walk forward. The spacebar and arrow keys do not depend on the layout, so in MIRA at first the only things that worked were jumping and driving with the arrows. Switch to an English layout before launching.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…