The hardest part of generating a 3D scene isn't creating it. It's changing it. Most AI systems treat a generated image as a static output, a flat picture. If you want to modify the left wall or swap out a character, you start over from scratch, losing all the coherent geometry you just paid for. Speridlabs, an AI lab based in Madrid, is building its entire bet on solving that problem first.
Its initial product, called Mundus, is described by the company as a “3D Midjourney” with a crucial twist: users can modify individual parts of a scene while the rest of the scene’s geometry stays intact [PERPLEXITY SONAR PRO BRIEF]. This focus on editable, persistent 3D generation is the technical wedge for a much broader ambition. The company is developing spatial foundation models designed to understand, generate within, and eventually reason over dynamic 3D worlds [Speridlabs, April 2026]. For now, it's targeting developers and creators in robotics, gaming, and special effects [TechCrunch, October 2026].
The wedge of editable geometry
The core technical claim is one of granular control. In a blog post outlining its vision, Speridlabs argues that existing systems from competitors like Runway, Odyssey, and Google's Genie cannot be queried or modified in the same localized way [PERPLEXITY SONAR PRO BRIEF]. For a game developer iterating on a level design or a robotics team simulating a warehouse pick-and-place task, this persistence matters. Regenerating an entire environment because one object is out of place is computationally expensive and breaks continuity. Mundus proposes a model that understands a scene as a structured, editable space, not just a collection of pixels.
This approach suggests a developer-first go-to-market. The company explicitly lists developers, researchers, and creators as its intended users [Speridlabs, April 2026]. An API or developer platform model would allow these technical users to integrate spatial reasoning into their own pipelines, whether for generating synthetic training data for robots or prototyping game assets.
Backing from a focused demo day
Speridlabs came out of stealth backed by Pear VC and Base10 Partners, though the specific round size and valuation remain undisclosed [Speridlabs, April 2026]. The company gained public visibility when it was featured among the startups presented at the PearX accelerator demo day in October 2026 [TechCrunch, October 2026]. PearX can provide startups with up to $2 million, but that figure is a program maximum and not a confirmation of Speridlabs' specific funding amount [Mezha, October 2026]. The backing from two established venture firms, however, signals belief in the technical team's approach to a foundational AI problem.
The founding team consists of brothers Chema and Guillermo Garabito [LinkedIn, retrieved 2026]. Chema Garabito, the CEO and Chief Scientist, publicly describes himself as a researcher, developer, and hacker [Chema Garabito’s public profile]. The available record shows a technical, product-building orientation rather than a purely academic research background.
The scale of the spatial bet
Speridlabs is entering a field crowded with well-resourced giants and ambitious startups, all chasing the promise of generative 3D. The company's differentiation rests on the specific capability of localized editing within a persistent scene. If it can deliver that reliably and at scale, it could carve out a defensible niche. The initial target sectors are logical first markets. Both gaming and visual effects have clear workflows for 3D asset creation, while robotics increasingly relies on high-fidelity simulation for training.
The technical breakdown here is about data efficiency and model architecture. Training a model to understand scene structure well enough to allow localized edits requires a different type of training data and likely a novel neural architecture compared to standard diffusion models. The promise is that once trained, the model is more useful and cost-effective for iterative design work. The risk is that achieving the necessary fidelity and generality for commercial use is a steeper research climb than anticipated.
What could go wrong at scale is a question of computational cost and market timing. Generating and, more importantly, persistently storing and modifying complex 3D scenes is computationally intensive. An API pricing model must account for these costs while remaining accessible to developers. Furthermore, the company is betting that the market's need for editable 3D generation will mature before larger players with more compute resources can replicate the functionality. The roadmap from a demo-day prototype to a robust, production-ready API serving enterprise robotics teams is long and fraught with engineering challenges beyond pure research.
Sources
- [Speridlabs, April 2026] The Shape of Intelligence | https://speridlabs.com/blog/01-post
- [TechCrunch, October 2026] 5 startups that caught VCs’ attention at the latest PearX demo day | https://techcrunch.com/2026/10/05/5-startups-that-caught-vcs-attention-at-the-latest-pearx-demo-day/
- [Mezha, October 2026] PearX Demo Day Spotlights 16 Startups Building AI, Robotics and Finance Tools | https://mezha.net/eng/news/14f1c4ac_pearx_demo_day/
- [Chema Garabito’s public profile] Chema Garabito's X profile | https://x.com/chemagarabito
- [LinkedIn, retrieved 2026] Speridlabs LinkedIn page | https://www.linkedin.com/company/speridlabs