Hardcore Reviews
Hardcore Reviews

Open Source Is Not a License to Sell: World Models Audited

This review skips image quality and runs four ledgers instead, on selecting among open world-model and spatial-intelligence projects (explicitly scoped apart from our image capability review, the batch-23 agent long-context cost ledger, and the batch-24 voice cloud-versus-local ledger): openness of code and weights, the commercial boundary of the license, the hardware bar, and data and control. Using 2026-09-14 GitHub API snapshots it compares seven options: Robbyant/lingbot-world-v2 (1,755 stars, CC BY-NC-SA 4.0, non-commercial; 8 GPUs for 14B and 2 for 1.3B per the repo), closed-source Genie 3 (public-report specs of 720p at 24fps, about one minute of consistency, no multiplayer), bilawalsidhu/gods-eye-view (32,399 stars, MIT, runs in the browser), Robbyant/lingbot-map (17,022 stars, Apache-2.0, ECCV 2026 oral, streaming 3D reconstruction), Robbyant/lingbot-world v1 (4,449 stars, Apache-2.0) and others. Three claims anchor it: open source does not mean commercially usable, since a non-commercial license is a hard boundary; these projects live on different layers, generative, reconstruction and front-end aggregation, so the first question is which link you need; and the hardware claims contradict each other across README examples, repo scripts and media, so budget from the official repository. Includes a comparison table and selection advice for three readers: individual researchers, small teams, and commercial integrators.

Published September 14, 20269 min read
<!-- open-world-model-comparison-review | review | Open Source Is Not a License to Sell: World Models Audited -->

This review makes one choice up front: we do not grade image quality. World models look their best in curated demo videos, which is exactly the part we trust least. Cost beats spectacle: a model earns follow-up for its real license, hardware, and data cost, not a staged minute. We state scope first, then show the table.

Evaluation Scope: Four Accounts, No Pretty Pictures

We fix the boundary to avoid overlapping existing comparisons. The site already published three with clear division of labor: image capability in reasoning-image-models-comparison-review, agent long-context cost in agent-long-context-cost-review, voice cloud-versus-local in cloud-vs-local-voice-ai-review. This piece does one job: take open-source world model apart and check what is truly open, what is commercially blocked, and what hardware you cannot afford. For the LingBot-World 2.0 hotspot see the LingBot-World 2.0 hotspot piece; for gods-eye-view resources see the gods-eye-view resources piece; for a local deploy SOP see the local deploy SOP piece.

We score four fixed accounts, in decision order, never reordered:

Account one, openness. Separate three things: are weights public, is code public, is deployment code public. Many projects dump weights on a hub and call themselves open, yet ship no inference script, no serving stack, no training code. That is downloadable closed source. A high star count does not mean you can reproduce.

Account two, the commercial boundary of the license. Licenses split three tiers: MIT and Apache-2.0 are commercially friendly, CC BY-NC-SA 4.0 explicitly bans commercial use, and between them sits fine print on attribution and reciprocal derivatives. The first step in selection is the license, not the star count.

Account three, hardware threshold. Every hardware number is tagged official or engineering estimate. What the README claims and what the repo script sets often differ, and press is usually more aggressive. When they conflict we show both and defer to the official repository.

Account four, data and control rights. Whose service runs the model, whose database holds your data, can you export, can you leave without lock-in: these are long-term costs. Closed subscriptions lose this account by design.

Every number comes from a GitHub API snapshot on 2026-09-14. Anything tagged reported is press or official-blog caliber, not our measurement. Prices we could not verify read subject to official, and we never invent a price.

The Four Accounts, Line by Line

Account one: openness (weights, code, deployment, kept separate)

LingBot-World 2.0 (repo Robbyant/lingbot-world-v2) publishes weights and some code for its 14B and 1.3B sizes, but deployment code is officially not open. You get the model and training code, not the production serving stack, so reproduction and redeployment cost you. Genie 3 is the opposite: fully closed, reachable only via a Gemini Ultra subscription, no weights. gods-eye-view (bilawalsidhu/gods-eye-view) is fully open code running locally in the browser, the highest openness here. lingbot-map and lingbot-world v1 are open-code Apache-2.0 projects with deployment code available. Split open into three layers and the seven projects fall into three openness levels; the label all open collapses.

Account two: the commercial boundary of the license (the core argument)

The point we most want clear: open source is not commercially usable. LingBot-World 2.0 ships CC BY-NC-SA 4.0. NC means NonCommercial, an explicit exclusion of any commercial purpose. SA means ShareAlike, forcing derivatives to the same license. Building a product, a paid service, or an internal commercial system on it crosses the line. The trap is that SA is contagious: outward improvements must also open under the same license, forcing private value public. gods-eye-view uses MIT: commercial use, closed redistribution, anything goes with the copyright notice kept. lingbot-map and lingbot-world v1 use Apache-2.0, commercially friendly but requiring attribution, a note on changes, and a default patent grant. Lay the tiers side by side and the first cut falls on the license, not stars.

Account three: hardware threshold (README, repo script, press often disagree)

The case is LingBot-World 2.0. The README example runs 1.3B on four cards; the same repo's run_fast.sh references two cards for 1.3B and eight for 14B. Both official, yet a factor of two apart. Press then writes runs on a consumer GPU flatly. For budgeting we recommend the repo script, the command you can reproduce; the README example is the optimistic version. Genie 3's 720p at 24fps is reported caliber; cost sits on Google's servers, so the user side has zero threshold but zero control. gods-eye-view runs locally on Node 24.x or 26.x, the lowest bar. lingbot-map does streaming 3D reconstruction, engineering estimate mid-to-high GPU; exact compute subject to official.

Account four: data and control rights

Genie 3 locks data and execution inside the Google subscription; ownership and terms of your output are set by the other party, the lowest control. LingBot-World 2.0 keeps deployment code closed, so private deployment means building your own stack; data sovereignty is yours but engineering cost is high. gods-eye-view aggregates public data (flights, ships, satellites, earthquakes, traffic, public cameras), binds to localhost, keeps data local, the best control and privacy. Open projects like lingbot-map and lingbot-world v1 keep data and control in your hands if you can deploy them. For compliance-bound enterprises this account often decides selection more than quality.

Side-by-Side Comparison

The table lists seven projects; OpenWorldLib and LightX2V are supplementary, all figures as of the 2026-09-14 snapshot. Any hardware or cost figure the official side did not price reads subject to official. Read the Type / Layer column first; it decides what the table compares.

ProjectType / LayerOpennessLicense and Commercial BoundaryHardware Threshold (official)Deployment Code OpenSnapshot
Robbyant/lingbot-world-v2 (LingBot-World 2.0)Generative world modelWeights + code publicCC BY-NC-SA 4.0, no commercial, derivatives must share1.3B ref 2 cards / 14B ref 8 cards (run_fast.sh)No2026-09-14
Genie 3 (Google DeepMind)Generative world modelClosedClosed subscription, commercial terms set by Google720p at 24fps (reported), zero user-side thresholdNo2026-09-14
bilawalsidhu/gods-eye-viewFront-end aggregation viewCode fully openMIT, commercially friendlyLocal browser, Node 24.x or 26.xYes2026-09-14
Robbyant/lingbot-mapReconstruction 3D / spatialCode openApache-2.0, commercially friendly (attribution)Streaming 3D reconstruction, est. mid-high GPUYes2026-09-14
Robbyant/lingbot-world (v1)Generative world model (v1)Code openApache-2.0, commercially friendly (attribution)Est. mid-high GPUYes2026-09-14
OpenDCAI/OpenWorldLib (supp.)World model toolkitCode openApache-2.0UnconfirmedYes2026-09-14
ModelTC/LightX2V (supp.)Video generation acceleratorCode openApache-2.0UnconfirmedYes2026-09-14

Read the table by Type / Layer first. Generative (LingBot-World 2.0, Genie 3, lingbot-world v1), reconstruction (lingbot-map), and front-end aggregation (gods-eye-view) are not the same problem domain. One table answers which link you need, not which is stronger. A digital-twin factory team and a real-time game team get different answers; one ruler yields the wrong conclusion.

Selection Advice by Reader Type

We give conclusions per reader type, not the line they are all good. Order: license, deployment openness, hardware, stars.

Individual researchers. Goal is reproducing a paper, running a demo, your own experiments. Prioritize openness and hardware. gods-eye-view under MIT and local browser is cheapest to play with and teach. For generative models, lingbot-world v1 under Apache-2.0 fits research derivatives better than v2 under CC BY-NC-SA 4.0, because v1 does not block the commercial off-ramp. The supplementary LightX2V suits video-generation acceleration. Caution: research is not commercial, but if you later ship a product, starting from v1 saves the rewrite that NC-locked v2 would force.

Small teams building a product. This is where the license red line hurts most. A paid or external service means LingBot-World 2.0 under CC BY-NC-SA 4.0 is out. gods-eye-view under MIT can be wrapped and closed safely. lingbot-map under Apache-2.0 is commercial, but keep attribution, note changes, read the patent clause. On hardware, gods-eye-view has near-zero marginal cost; lingbot-map's threshold needs a confirmed GPU budget. Worst habit: ship first, fix compliance later; an NC license used commercially is infringement, not a late notice away.

Commercial integration (enterprise or vendor). Control matters most. Closed Genie 3 fits a validate-demand prototype, but long-term integration must price locked data and mutable terms; a price rise can break delivery. What you can self-host and audit are the Apache-2.0 or MIT open projects. Advice: build the skeleton on open projects, write the boundary into procurement and legal review, then add a closed subscription if needed. For regulated industries, open with deployment code is a hard gate; Genie 3 is off the list by root.

Cold Take

First, stars are not usability. gods-eye-view's thirty-thousand-plus stars are partly traffic from a cool front end over public data, solving presentation not generation. Comparing it with generative models on who is the world model compares the wrong axis. Stars lag attention, not lead usability.

Second, the zero threshold of closed subscriptions is sugar coating. Genie 3 costs nothing up front, but the price is zero control, subscription billing, mutable terms. For product teams this is a time bomb: a demo that runs today may not run tomorrow. Betting core capability on a subscription hands over your lifeline.

Third, officially open and deployable open are separated by a river. LingBot-World 2.0 opens weights and code but not deployment code, shifting reproduction cost to you. When judging openness, always ask whether deployment code is open; that question decides if you ship in three months.

Fourth, hardware-caliber inflation is easiest to miss. When README, repo script, and press disagree, press is most optimistic, README next, repo script most conservative. Budget against the reproducible command, defer to official when in doubt.

Fifth, hype is not maturity. Generation, reconstruction, aggregation sit at different maturity: generative is flashiest but tightest on license and heaviest on engineering; reconstruction is practical but far from world; aggregation is lightest but not generation. Stirring them into one concept is the biggest water in this hype.

One line to close: the takeaway is layering. Generation, reconstruction, aggregation are three jobs; pick the wrong layer and no model saves your requirement.

FAQ

Q1: How do open-source licenses actually split the commercial boundary?

A1: Three tiers. MIT is widest: commercial, closed, redistribute, just keep the copyright notice. Apache-2.0 is commercially friendly but requires attribution, a note on changes, and carries patent grant terms. CC BY-NC-SA 4.0 bans commercial use via NC and forces reciprocal derivatives via SA, which eliminates product use. Step one is the license, not stars.

Q2: Can I use LingBot-World 2.0 commercially?

A2: No, not directly. It ships CC BY-NC-SA 4.0; NonCommercial excludes commercial purpose and ShareAlike forces the derivative to the same license. Paid services, external products, internal commercial systems all cross the line. For commercial generative models, move to lingbot-world v1 under Apache-2.0 or gods-eye-view under MIT.

Q3: Which hardware number should I trust?

A3: Trust the reproducible repo script; it is usually more conservative than the README example and press. LingBot-World 2.0 is the case: README runs 1.3B on four cards, run_fast.sh references two for 1.3B and eight for 14B. Budget against the script; when conflicted, defer to the official repository, treat press as reference only.

Q4: Does gods-eye-view count as a world model?

A4: No, it is a front-end aggregation spatial view, not generative. It aggregates public data (flights, ships, satellites, earthquakes, traffic, public cameras) locally in the browser, solving see the real world in real time, not generate a virtual world. Against generative models it answers which link you need, not which is stronger.

Q5: What is the first step in selection?

A5: First check whether the license covers your use, then whether deployment code is open, and only last look at stars and demos. Stars are traffic; the license is the line between life and death. Commercial projects must exclude CC licenses first, then talk models and hardware.

This article is AI-assisted and human-edited. Last updated: 2026-09-14

FAQ

How do open-source licenses actually split the commercial boundary?
Three tiers. MIT is widest: commercial, closed, redistribute, just keep the copyright notice. Apache-2.0 is commercially friendly but requires attribution, a note on changes, and carries patent grant terms. CC BY-NC-SA 4.0 bans commercial use via NC and forces reciprocal derivatives via SA, which eliminates product use. Step one is the license, not stars.
Can I use LingBot-World 2.0 commercially?
No, not directly. It ships CC BY-NC-SA 4.0; NonCommercial excludes commercial purpose and ShareAlike forces the derivative to the same license. Paid services, external products, internal commercial systems all cross the line. For commercial generative models, move to lingbot-world v1 under Apache-2.0 or gods-eye-view under MIT.
Which hardware number should I trust?
Trust the reproducible repo script; it is usually more conservative than the README example and press. LingBot-World 2.0 is the case: README runs 1.3B on four cards, run_fast.sh references two for 1.3B and eight for 14B. Budget against the script; when conflicted, defer to the official repository, treat press as reference only.
Does gods-eye-view count as a world model?
No, it is a front-end aggregation spatial view, not generative. It aggregates public data (flights, ships, satellites, earthquakes, traffic, public cameras) locally in the browser, solving see the real world in real time, not generate a virtual world. Against generative models it answers which link you need, not which is stronger.
What is the first step in selection?
First check whether the license covers your use, then whether deployment code is open, and only last look at stars and demos. Stars are traffic; the license is the line between life and death. Commercial projects must exclude CC licenses first, then talk models and hardware.

Related

Hardcore Reviews

Cloud vs Local Voice AI: A Cost and Control Showdown

This review ignores capability and runs the cost-and-control numbers on two routes for voice AI: the cloud real-time speech API versus local open-source tooling (explicitly scoped apart from our 8-26 image-model capability review, batch-22 image cost ledger, and batch-23 agent long-context cost ledger). It opens by arguing voice cost is harder to model than text or images: real-time behavior, concurrency, duration distribution, language/dialect coverage, and privacy compliance stack five dimensions at once. It then compares the two routes dimension by dimension - unit price and billing, latency and real-time, privacy/compliance, controllability/customization, language coverage - pitting cloud representative GPT-Live-1 (closed-source, metered, real-time out of the box) against local representative VoiceStudio (open-source, one-time compute, data stays local, engines swappable), with a five-dimension scorecard and a five-scenario selection table, and concludes by scale: individual, small team, bulk. Every unit price is symbolic (P_cloud / C_local) or marked "per official pricing page"; magnitude judgments are engineering estimates. Cold take: a vendor's "pay-as-you-go saves" only covers the one workload inside its chosen sweet spot; concurrency N, duration distribution T, and mandatory real-time decide the actual bill, so measure yourself.

Sep 13, 20269 min read
Hardcore Reviews

Agent Context Cost Review: Accounting for Every Tool Call

This review ignores capability and runs the numbers instead, on the context cost of agentic long-context and multi-turn trajectories (explicitly scoped apart from our 8-26 image-model capability review and the batch-22 image cost piece). It opens with a reproducible per-turn cost formula and makes the point that the resident prefix is the portion you re-pay on every single turn. It then compares five levers - prefix caching, KV Cache compression and sparse attention, context compression, tool-output trimming, and switching trajectory replay to incremental commit - across payoff magnitude, implementation cost, risk and fit, with a five-lever comparison table plus a cost-structure table for three scenarios (a ten-tool-call single task, a long-trajectory coding agent, and batch offline work), then ranks the levers by scale: individual, small team, and bulk. Every unit price is either symbolic (P_in / P_out / P_cache) or marked "refer to the official pricing page"; magnitude judgments are labeled engineering estimates, never passed off as benchmarks. Cold take: a vendor's "cost down X%" is usually the optimum under one specific workload - cache hit rate, context distribution and tool-output length decide your bill, so instrument your own stack rather than trusting launch numbers.

Sep 10, 20269 min read
Hardcore Reviews

Closed API vs Open Weights: What Does One Image Really Cost

With ChatGPT Images 2.5 and Ant's open-source LLaDA-Image landing in the same week, text-to-image has split into closed APIs versus self-hosted open weights. This review ignores image quality and runs the cost-and-control numbers instead: five routes - closed APIs, self-hosted open weights, per-second third-party inference platforms, local consumer hardware, and domestic cloud APIs - with per-image cost projected at two volumes (100 and 10,000 images per day), plus a comparison table and scenario-based selection (hobby use, e-commerce batch, data-sensitive industries, brand-style fine-tuning, maximum quality). It flags four traps: undeclared licenses, cold starts on per-second billing, Chinese text rendering, and cross-border data transfer. Explicitly scoped apart from our 8-26 capability review of reasoning image models. Representative comparison, not hands-on benchmarking; pricing per official sites.

Sep 9, 20269 min read