Hardcore Reviews
Hardcore Reviews

Real-Time World Models Compared: Playable or Just Fast?

A comparison of world models along one axis: from generating to being playable, accounting only for real-timeness, the interaction loop, hardware floor and openness. It does not redo this site's September 14, 2026 four-account comparison, nor retell the July LingBot-World 2.0 launch or its local-deploy SOP. It opens by setting the ruler: real time means at least three different things - model processing throughput, end-to-end action-to-photon latency, and browser presentation frame rate - and they can differ by orders of magnitude, so comparing them directly guarantees a distorted answer; the best evidence comes from Matrix-Game 4.0's own README, which labels --fps 40 a target capture cadence rather than a throughput or latency guarantee. Seven contenders on one table, all star counts from a GitHub API snapshot on 2026-10-16 and stated as fact, not as quality: Matrix-Game 4.0 (Skywork, Kunlun Wanwei) at 2,488 stars, code Apache-2.0 per the v4 directory plus MIT at the root, officially 480p 60 FPS and sub-150 ms latency not independently reproduced, README-measured at about 38-39 FPS on an A100 at 864 x 480, hardware A100 80 GB or RTX 5090; MineWorld (Microsoft) at 494 stars, MIT, Diagonal Decoding at 4-7 FPS with a claim of real-time interaction in the Minecraft context, developed and tested on A100 and H100, and with its Hugging Face checkpoints temporally taken down since May 2025; HY-World 1.5 / WorldPlay (Tencent Hunyuan) at 1,619 stars with an undeclared NOASSERTION license that must be read from the repository, README stating 24 FPS with a next chunk of 16 frames and VRAM of 28 GB at sp=8, 34 GB at sp=4 and 72 GB at sp=1; HunyuanWorld-Voyager (Tencent Hunyuan) at 1,604 stars, also undeclared, and explicitly not real-time - 540p needs at least 60 GB and 80 GB is recommended, and 512 x 768 at 49 frames and 50 steps takes 1,925 seconds on one H20 and 288 seconds on eight; Helios (PKU-YuanGroup) at 2,206 stars, Apache-2.0, a 14B real-time long-video model officially at 19.5 FPS end-to-end on a single H100 without the usual acceleration tricks, with a user-measured 20.89 FPS, about 10 FPS on Ascend NPU, autoregressive in 33-frame chunks, and an interactive inference script still marked as under development; and two closed references, PixVerse R1 (AI-SHITECH, launched January 2026, officially the first universal 1080P real-time world model with an instantaneous response engine compressing 50-plus sampling steps to 1-4) and Genie 3 (Google DeepMind, reported at 720p 24fps with about a minute of consistency and no multiplayer). The verdict comes in three tiers: for self-hosting and code changes, check the license then the VRAM; for experience without deployment, go cloud and closed; and for long-horizon exploration and reconstruction on your own hardware, budget 60-80 GB. The core judgment is that plenty of projects can generate and very few can be played, with playability locked by three keys at once - definitions, VRAM and licensing - and it warns that Helios's group offloading, which can cut VRAM to about 6 GB, trades throughput for memory rather than proving anything runs playably on 6 GB.

Published October 16, 202610 min read
<!-- realtime-playable-world-models-comparison-review | review | Real-Time World Models Compared: Playable or Just Fast? -->

Every world-model demo looks more spectacular than the last, yet generating a video and playing one are different things. Generating means turning a prompt into a convincing clip. Playing means you press a key and the next frame reacts, with a delay short enough that your brain never notices the break. This piece is not about image quality, and it does not re-run the selection math. It asks one question: as of October 2026, how many open projects have crossed the threshold where you can actually play them.

1. Setting the Ruler: Three Meanings of "Real Time"

The common mistake is flattening everyone's numbers onto one line. But "real time" refers, in the source documents, to at least three different quantities. The first is model processing throughput, how many frames a model chews per second. The second is end-to-end latency, the time from an input action to a visible frame, usually called action-to-photon. The third is presentation frame rate, the smoothness a user actually sees. These can differ by orders of magnitude, and mixing them guarantees a false conclusion.

The cleanest evidence comes from the v4 README of Matrix-Game 4.0, which draws the line itself. It labels --fps 40 as a target capture cadence, not a throughput or latency guarantee. Its A100 numbers are labelled as model processing throughput inside an online session, plus server-side pipeline throughput that includes waiting for game frames, and the README adds that neither is browser presentation FPS or action-to-photon latency. So the rule here is to separate the measurement basis first, then compare, marking every vendor figure as not independently reproduced.

2. Seven Contenders on One Table

Our 2026-09-14 comparison, "Open Does Not Mean Commercial: Four Ledgers for World Models" (open-world-model-comparison-review), tallied four ledgers, meaning license, hardware, data and openness, across seven entries including lingbot and gods-eye. This piece does not repeat that selection math; it tallies only the playability layer. Our July features on LingBot-World 2.0 and the small-model local deployment guide cover a single release and installation steps respectively, so this article restates neither. All star counts below are GitHub API snapshots from 2026-10-16, recorded as fact, not as a quality verdict.

ProjectStarsLicenseReal-time / playable figureHardware floor
Matrix-Game 4.0 (Kunlun Wanwei / Skywork)2,488Code Apache-2.0 (v4 directory text) plus MIT at rootVendor: 480p 60FPS, latency under 150ms; README self-test: about 38-39 FPS at 864x480 on A100A100 80GB / RTX 5090
MineWorld (Microsoft)494MITDiagonal Decoding at 4-7 FPS, described as real-time interactiveA100 / H100
HY-World 1.5 / WorldPlay (Tencent Hunyuan)1,619Undeclared (defer to the repo text)README states 24 FPS; next chunk equals 16 frames28G to 72G by parallelism
HunyuanWorld-Voyager (Tencent Hunyuan)1,604Undeclared (defer to the repo text)Not real time: 540p needs at least 60GB of VRAM60-80GB
Helios (PKU-YuanGroup)2,206Apache-2.019.5 FPS on a single H100; a user measured 20.89 FPSSingle H100
PixVerse R1 (AIsphere, closed reference)n/aClosedVendor: 1080P general model, real-time positioning, 1-4 stepsCloud
Genie 3 (Google DeepMind, closed reference)n/aClosedPublic reporting: 720p at 24fps, about one minute of consistencyCloud

3. The Five Open Models and Two Closed References

Matrix-Game 4.0 (Skywork): one project, three measurement bases. The repository is SkyworkAI/Matrix-Game, address: github.com/SkyworkAI/Matrix-Game. The license chain reads piece by piece: the root LICENSE is MIT, while Matrix-Game-4/LICENSE in the v4 directory is the full Apache-2.0 text, and the v4 README states that project code is Apache-2.0 while vendored components keep their own licenses; the bundled TAEHV component is MIT. Three performance versions coexist. The vendor figure is a single RTX 5090, 480p at 60FPS, end-to-end latency under 150ms, 83 FPS steady state and 18.5 GiB peak VRAM, and it is not independently reproduced. The official performance panel is still a placeholder reading 0FPS, 0ms, 0p and 0GPU, so the slogan numbers have not landed yet. The hardest version is the README self-test: one A100-SXM4-80GB at 864x480, shooters at about 38-39 FPS fixed-prompt as model processing throughput, and racing at 38.5 fixed and 28.9 dynamic as server-side pipeline throughput from an earlier checkpoint of the same architecture. That 38-39 is throughput, not browser FPS. The 10-15 RTX 5090 presets use FP8 weights. Weights and assets defer to the Hugging Face model card, which our proxy blocked, so no license is guessed here.

MineWorld (Microsoft): the honest 4-7 FPS. The repository is microsoft/mineworld, address: github.com/microsoft/mineworld. Its core trick is Diagonal Decoding, and its stated figure is a generation rate of 4 to 7 frames per second, described as enabling real-time interaction in open-ended game environments. That real-time lives in the Minecraft context, where 4-7 FPS is enough to interact, and it should not be compared head-on with a 60FPS slogan. One fact must be recorded: since May 2025 the Hugging Face checkpoints have been temporarily taken down, so it reads more like a reproducible research artifact than a ready product.

HY-World 1.5 / WorldPlay (Tencent Hunyuan): 24 FPS, sixteen frames at a time. The repository is Tencent-Hunyuan/HY-WorldPlay, address: github.com/Tencent-Hunyuan/HY-WorldPlay. The GitHub API flags NOASSERTION and no specific license is named, so the repository declares no license and the repository LICENSE text governs, with commercial risk on the user. Its introduction states that it generates long-horizon streaming video at 24 FPS, and inference advances 16 video frames per next-chunk prediction before Reconstituted Context Memory rebuilds context from past frames. Its VRAM accounting is detailed: with the autoregressive distilled model at 125 frames, sp=8 costs 28G, sp=4 costs 34G and sp=1 costs 72G. It marks itself real-time in its own evaluation table, and it open-sources data, training, RL post-training and distillation.

HunyuanWorld-Voyager (Tencent Hunyuan): explicitly not real time, and the steepest floor. The repository is Tencent-Hunyuan/HunyuanWorld-Voyager, address: github.com/Tencent-Hunyuan/HunyuanWorld-Voyager. Again NOASSERTION: the repository declares no license and the repository LICENSE text governs. Its requirements table says 540p, batch size one, 60GB peak GPU memory, with 60GB minimum and 80GB recommended, and its output is camera-controllable RGB-D world video plus an exportable point cloud, not a frame-by-frame response to keyboard and mouse. The latency table is clearer still: at 512x768, 49 frames and 50 steps, one H20 takes 1,925 seconds and eight H20s bring that only to 288 seconds. It belongs here as a reverse ruler: the same four words, world model, but its time scale is minutes to half an hour, an order of magnitude away from play.

Helios (PKU-YuanGroup): 19.5 FPS on one H100. The repository is PKU-YuanGroup/Helios, address: github.com/PKU-YuanGroup/Helios. Its Apache-2.0 license is the cleanest among the five. It is a 14B real-time long-video model whose vendor figure is 19.5 FPS end-to-end on a single H100, without the usual long-video anti-drift strategies and without standard acceleration such as KV-cache, causal masking, sparse or linear attention, TinyVAE or quantization. On an Ascend NPU it reaches about 10 FPS, and one user measured 20.89 FPS on better hardware. It generates autoregressively in chunks of 33 frames. Its key caveat: the interactive inference script states the feature is still under development and results may not always meet expectations, so 19.5 FPS means high-throughput long-video generation, not a game-grade loop. Group offloading can push VRAM to about 6GB, but that trades throughput for memory and is not the same as playable at 6GB.

Two closed references. PixVerse R1 (AIsphere) is closed; the vendor describes it as the world's first 1080P general real-time world model, announced January 2026 and invite-only, with an instant-response engine that compresses more than 50 sampling steps to 1-4. Those are vendor and press figures, not independently reproduced. Genie 3 (Google DeepMind) is also closed, with public reporting of 720p at 24fps, roughly one minute of consistency and no multiplayer, a basis we also used in our September 14 comparison, so it counts as second-hand. Neither offers weights or a VRAM figure: at this layer, real time is rented from the cloud, never self-hosted.

4. Hardware: From One Consumer Card to 80GB of VRAM

Sorted by hardware, the spread is extreme. At the consumer single-card tier, Matrix-Game 4.0 is the only entry with a dedicated RTX 5090 preset, and 5080 and 5070 arrive next week. Above it: the single-H100 tier (Helios at 19.5 FPS, long video, interactive mode in development), then the 60-80GB tier (Voyager needs 60GB at 540p and is not real time; HY-WorldPlay needs 72G at sp=1). At the top is the closed cloud tier: PixVerse R1 and Genie 3 need no deployment, at the cost of never being self-hosted. One trap: Helios can compress VRAM to about 6GB through group offloading, but that trades throughput for memory, so "runs in 6GB" is not "plays in 6GB".

5. Openness: Apache-2.0, MIT, Undeclared, Closed

Split by license, the boundary is clear. The first tier is Apache-2.0: the v4 code of Matrix-Game 4.0 and Helios, commercial-friendly, requiring that notices be kept and modifications noted, and granting a patent license. The second tier is MIT: MineWorld, the older code at the root of Matrix-Game, and the bundled TAEHV component. The third tier is undeclared: HY-WorldPlay and Voyager, flagged NOASSERTION on GitHub, so the repository declares no license and the repository LICENSE text governs, with commercial risk on the user. The fourth tier is closed: PixVerse R1 and Genie 3. A license decides whether you may use something, not whether it is any good, and star counts are snapshots, not verdicts.

6. What to Pick, by Need

Tier one, self-hosting and source modification. Check the license first, then the VRAM. For Apache-2.0 plus a consumer card, Matrix-Game 4.0 satisfies both; for 14B-class throughput under Apache-2.0, look at Helios and accept that interactive mode is still in development. Tier two, experience only: PixVerse R1 and Genie 3 work out of the box, at the cost of never being self-hosted and mostly vendor or second-hand figures. Tier three, long-horizon exploration and reconstruction: Voyager and HY-WorldPlay fit better, but budget for 60-80GB of VRAM and accept a time scale that is not frame-by-frame playable. For pure research reproduction, MineWorld is honest, light and readable. In one line: many projects can generate, few can be played, and playability is locked behind three gates, the measurement basis, the VRAM and the license.

FAQ

Q1: How do the three meanings of "real time" differ?

A1: They are model processing throughput, end-to-end latency and presentation frame rate. Throughput is how many frames a model processes per second. Latency is the time from an input action to a visible frame, the action-to-photon figure. Presentation frame rate is what a user sees. The Matrix-Game 4.0 README separates all three and states that --fps 40 is a capture cadence target, not a throughput or latency guarantee.

Q2: Can the 480p 60FPS figure for Matrix-Game 4.0 be trusted?

A2: Read it in layers. The vendor figure of 480p 60FPS, latency under 150ms and 18.5 GiB peak VRAM is a vendor claim, not independently reproduced, and the official performance panel is still a placeholder. The README self-test reports about 38-39 FPS at 864x480 on an A100, and that is model processing throughput, not browser presentation FPS.

Q3: Can I use a NOASSERTION project commercially?

A3: Not by assumption. HY-WorldPlay and Voyager are flagged NOASSERTION, and no specific license is found. The accurate wording is that the repository declares no license and the LICENSE text governs, and before commercial use you must read the LICENSE text, the weights license and the asset license. If a license is not stated, the risk is yours.

Q4: With only one consumer-grade card, what can I run?

A4: On the current snapshot, the one entry with an explicit consumer preset is Matrix-Game 4.0, which ships an RTX 5090 preset and says 5080 and 5070 arrive next week. Helios can push VRAM down to about 6GB through group offloading, but that trades memory for throughput. The rest need a single H100, or 60-80GB of VRAM, or exist only in the cloud.

Q5: Do more stars mean more playable?

A5: No. Stars are a 2026-10-16 snapshot, not a playability score. Helios has impressive throughput but an interactive mode still in development; Voyager is explicitly not real time; MineWorld is the one that honestly states a 4-7 FPS interactive figure.

Closing Thoughts

The takeaway is not whose frame rate is highest, but that "real time" points at different things in different projects. Generating is no longer hard. Making throughput, latency and presentation all clear the bar of play at the same time is hard, and that bar sits between two walls, one of license and one of VRAM. Decide first which kind of real time you need, then check the license and the VRAM, and only then look at frame rate.

This article is AI-assisted and human-edited. Last updated: 2026-10-16

FAQ

How do the three meanings of "real time" differ?
They are model processing throughput, end-to-end latency and presentation frame rate. Throughput is how many frames a model processes per second. Latency is the time from an input action to a visible frame, the action-to-photon figure. Presentation frame rate is what a user sees. The Matrix-Game 4.0 README separates all three and states that `--fps 40` is a capture cadence target, not a throughput or latency guarantee.
Can the 480p 60FPS figure for Matrix-Game 4.0 be trusted?
Read it in layers. The vendor figure of 480p 60FPS, latency under 150ms and 18.5 GiB peak VRAM is a vendor claim, not independently reproduced, and the official performance panel is still a placeholder. The README self-test reports about 38-39 FPS at 864x480 on an A100, and that is model processing throughput, not browser presentation FPS.
Can I use a NOASSERTION project commercially?
Not by assumption. HY-WorldPlay and Voyager are flagged NOASSERTION, and no specific license is found in their repositories. The correct wording is that the repository declares no license and the repository LICENSE text governs, and before any commercial use you must read the LICENSE text, the weights license and the asset license word by word. If a license is not stated, the risk is yours.
With only one consumer-grade card, what can I run?
On the current snapshot, the one entry with an explicit consumer preset is Matrix-Game 4.0, which ships an RTX 5090 preset and says 5080 and 5070 arrive next week. Helios can push VRAM down to about 6GB through group offloading, but that trades memory for throughput. The rest need a single H100, or 60-80GB of VRAM, or exist only in the cloud.
Do more stars mean more playable?
No. Stars are a 2026-10-16 snapshot, not a playability score. Helios at 2,206 stars has impressive throughput but an interactive mode still in development. Voyager at 1,604 stars is explicitly not real time. MineWorld at 494 stars is the one that honestly states a 4-7 FPS interactive figure.

Related

Hardcore Reviews

Cheaper Per Token, Same Bill: LLM API Pricing Compared

Per-token list prices keep falling, but your bill may not: this comparison unpacks the unit-price illusion in LLM APIs - verbose output (Artificial Analysis measured K3.1 Pro at 1.7x the median token spend), cache-hit vs miss gaps up to 50x (DeepSeek 0.0028 vs 0.14 dollars), about 5.5 percent aggregator markups, and assorted surcharges. Includes a multi-source pricing table and workload-based picks, with a reminder that self-hosting moves cost rather than erasing it.

Nov 4, 20269 min read
Hardcore Reviews

Eight Flagship LLMs, Only Half Accept Your Image

A November 4, 2026 snapshot of the OpenRouter API lists 453 models, 308 of them accepting image input. This comparison puts eight flagships side by side on the same API fields: input modality, context length, and per-million-token input/output/cache-read prices. tencent/hy4 is text+image with a 1,048,576 context at 0.417/1.2505/0.021 dollars, while tencent/hy4-preview, GLM-4.6, Kimi K2 Thinking and DeepSeek V3.2 remain text-only on that channel. Includes a full table, scenario picks, and three caveats: no cross-leaderboard scoring, aggregator prices are not vendor list prices, and modality tags are not capability measurements.

Nov 4, 20269 min read
Hardcore Reviews

Your First 3D Model: Six AI Modeling Tools Tested

From zero to your first usable 3D model: a six-tool comparison of Kimi 3D Make (script-driven parametric modeling), Meshy (full-pipeline all-rounder), Tripo (speed king), Rodin (studio-grade style tuning), plus free self-hosted Tencent Hunyuan3D and TRELLIS 2. Covers route differences, ease of use, free tiers, pricing, and main weaknesses, with a master table, scenario picks, and three beginner pitfalls (topology cleanup, marketing-shot gap, organic shapes).

Nov 3, 20269 min read