Meta AI Hackathon 2026 Results: Winners, Prizes and Projects
Official organizer update (November 10, 2026): Cerebral Valley announced on Discord that 5,367 projects were submitted and published individual gallery links for all 14 named prize recipients. The rankings below now link to those project pages. The $10,000 Community Vote remains open, with its winner due November 16. See the official winner announcements and full closing show.
Updated November 10, 2026. The Meta Global AI Developer Hackathon, hosted by Meta and Cerebral Valley, awarded ten main placements across Startup and Indie Developer tracks and four named bonus prizes during its November 9 ceremony. The host reported approximately 5,367 projects. The advertised prize pool was $1 million: $950,000 in main awards and $50,000 in bonuses, including a $10,000 Community Vote prize whose winner was scheduled for November 16.
The bigger story: winning projects move beyond chatbots toward AI that observes real environments, performs multi-step work and checks its outputs. From construction inspections to software security, the strongest demonstrations made the result inspectable rather than merely plausible.
Results at a glance
Startup winners
| Rank | Project | Prize |
|---|---|---|
| 1 | Consai — Punch Walk | $250,000 |
| 2 | Formwright | $150,000 |
| 3 | ORCA by IssueLab | $75,000 |
| 4 | Voluma Bio | $50,000 |
| 5 | Kroonen AI — GlimmerScope | $25,000 |
Indie Developer winners
| Rank | Project | Prize |
|---|---|---|
| 1 | ControlAnything | $150,000 |
| 2 | Gatekeeper | $100,000 |
| 3 | David Barry — Basement Ready | $75,000 |
| 4 | Aurora — Nemesis | $50,000 |
| 5 | Video2URDF | $25,000 |
Startup projects: what they built and why the judges cared
Consai — Punch Walk: an inspection becomes a verified workflow
The $250,000 Startup winner turns spoken, filmed construction walkthroughs into defect lists sorted by trade, room and severity. Each task links back to the inspector’s words and a precise video moment; a second walkthrough can compare reported repairs with earlier evidence. In the demo, an incorrectly claimed repair was flagged for human review. The team reported detecting 12 of 12 defects in its labeled example and converting a 105-second walkthrough into a report in under five minutes for around $0.54. These are presenter-reported demonstration results, not an independent field accuracy study. Judges emphasized the clear customer problem, practical workflow and human oversight.
Formwright: from words to manufacturable electronics designs
Formwright translates a description of a custom electronic device into schematics and circuit-board designs intended for manufacturing and ordering. Judges saw a potential market for personalized hardware: rather than search for a niche product, describe what you need. That vision still depends on electrical checks, component availability, manufacturing validation and real-world reliability. The demo is a promising prototype, not evidence that arbitrary hardware can be produced safely on demand.
ORCA by IssueLab: coordinate 3D design changes
ORCA addresses clashes between engineering or building models. Its demonstration described detecting 464 conflicts in example models, including a duct intersecting a structural beam, then classifying issues, proposing edits and applying changes after user approval. IFC editing, BCF export, STEP files and reporting were shown. The 464 count belongs to that demo, not a universal performance benchmark. Its value proposition is coordinating and checking design changes, not merely generating 3D geometry. Project repository: IssueLab-Inc/ORCA.
Voluma Bio: research measurements with review checkpoints
Voluma Bio presented analysis of micro-CT bone scans relevant to preclinical osteoarthritis research. A Muse Wonder orchestrator delegates work to specialist agents; SAM 3.1 helps segment structures such as the femur and tibia. The system exposes uncertain measurements for researcher review. Judges valued that feedback loop. This is research assistance, not a clinically validated diagnostic system.
Kroonen AI — GlimmerScope: steer a model from inside
GlimmerScope visualizes internal model activations and tests interventions on concepts, using sparse autoencoders across nine layers. The creator reported more than 519,000 labeled features and over 170,000 steering tests. Judges praised its research depth and accessible visualization but questioned the immediate startup business model. Steering a feature is not the same as guaranteed control over model behavior.
Indie projects: interfaces, security, housing and robotics
ControlAnything: spatial controls built on request
The $150,000 Indie winner uses Meta Quest 3, hand tracking and Muse agents to generate controls for objects and software. Its demo showed a computer, app interactions, a USB MIDI synthesizer and other connected hardware. Judges highlighted accessibility and the compelling “Iron Man” interface. The demonstration does not establish compatibility with every device or production-grade safety.
Gatekeeper: investigate npm releases before installation
The $100,000 runner-up monitors newly published npm packages using static inspection without executing unknown code, agentic investigation and a fine-tuned Llama Guard model. Its creator reported 48 releases later confirmed as malware and a median 14-hour early warning. Those results have not been independently reproduced. The important idea is a defensive checkpoint in the software supply chain, not a universal Windows-update antivirus.
David Barry — Basement Ready: preliminary renovation planning
Basement Ready analyzes phone footage of a basement to estimate layout and assist with preliminary conversion planning against local building requirements. It uses Meta perception and reasoning tools, allows measurement corrections and produces draft drawings and forms. Final permitting remains with the local authority. Project: basementready.com.
Aurora — Nemesis: train robots against simulated hazards
Nemesis creates simulation challenges for quadruped and humanoid robots, uses reinforcement learning to retrain policies and checks regressions before deployment. It can derive scene elements from imagery. Judges were interested in how simulated improvements transfer to real robots. Demo: nemesis.divagat.com.
Video2URDF: articulated robotic models from footage
Video2URDF uses phone video and visual models to estimate an object’s geometry and joints, producing descriptions for robotics simulation. The demo covered items such as drawers and a microwave door. The jury raised a crucial question: which dimensions are measured and which are inferred? That distinction determines whether generated models are reliable enough for downstream use.
Bonus awards and the pending community vote
| Award | Winner | Prize |
|---|---|---|
| Most Viral Project | Adrian — Playtest Agent | $20,000 |
| Best Build on Open Weights | Tessa | $10,000 |
| Best Agentic Use-Case | oqed | $5,000 |
| Best Creative Build | Synthexis | $5,000 |
| Community Vote | Not announced at the ceremony; scheduled November 16 | $10,000 |
Adrian demonstrated a screenshot-driven game-testing agent that reports bugs and attempts reproducible replays. The host cited around eight to nine million X views and acknowledged paid promotion; that reach is not independently verified here. Tessa reported screening approximately 76,000 stars and surfacing eight exoplanet candidates, not confirmed planets; the creator described 92% precision on an internal evaluation. oqed proposed faster assembly for C/C++ hot paths, rejecting candidates that fail compilation, tests or performance checks; its claimed 27–32% speedup was workload-specific. Synthexis used gestures and natural language to map hand movements to audio-plugin controls.
Four trends that explain the results
1. AI that finishes a task
Consai turns a site visit into actionable work orders; ORCA moves from a collision to a proposed model edit; Formwright moves from an idea to an electronics design. These are workflows with inputs, outputs and follow-through, not just chat responses.
2. Verification becomes part of the product
Gatekeeper checks suspicious packages, oqed tests optimized code, Consai retains repair evidence, and Voluma Bio flags uncertain measurements. “Verified” means different things in each domain: passing tests is not a universal mathematical proof, and photographic evidence is not a safety certificate. But the direction is unmistakable: useful AI must make results reviewable.
3. Natural interfaces meet physical constraints
ControlAnything and Synthexis replace some menus with gestures; Formwright and Nemesis bring AI closer to hardware and robotics. That raises practical requirements around latency, permissions, tolerances and safety that a polished demo cannot settle on its own.
4. Specialized research and clear commercial value are different strengths
Tessa, Voluma Bio and GlimmerScope illustrate the promise of domain-specific AI. Judges celebrated GlimmerScope’s sophistication while questioning its business model. By contrast, Consai won the Startup category with a problem, customer and operational benefit that were immediately legible. This is an editorial interpretation of public feedback, not a claim about private scoring.
What happens next?
The strongest signal from the hackathon is not that AI can do everything autonomously. It is that focused systems can observe, act and leave evidence for people to check. Whether these prototypes become durable products depends on real-world evaluation, integration, costs, safety and human accountability. The next competitive advantage may be not simply generating an answer, but completing a job whose result can be trusted and checked.
Judging transparency: participants seek answers after the awards
Alongside congratulations for the winners, some hackathon participants have expressed frustration and asked for greater clarity about how 5,367 submissions were screened and judged. In the event's Discord community, builders behind projects including TaskLens and SnapPart asked whether non-finalists could receive individual feedback, scores, ranking information or a way to challenge a possible evaluation error. Another participant asked whether those outside the top projects could learn where they placed. These are questions raised by individual participants, not evidence of widespread dissatisfaction among all entrants.
What the organizers have clarified
Some developers initially worried that their demos had not been reviewed because their public YouTube videos showed few or no views. Natasha, identified in Discord as part of Cerebral Valley, explained that demo videos were pulled into an internal platform for viewing and review. Public YouTube view counts therefore cannot establish whether a submission was watched or assessed. That clarification addresses an important source of concern, although it does not establish how much review each individual submission received.
The published judging rubric assigned 30% to Impact, 25% to Use of Meta AI Tooling, 20% to Technical Depth and Execution, and 25% to Demo. In responses to participant questions, the Cerebral Valley Discord assistant said organizers had not announced individual judging feedback or score breakdowns, full rankings, shortlist counts, details of any AI-assisted judging, or a process for confirming that a specific submission had been processed. It also said no review or appeal route had been announced. These statements describe what had been publicly announced at the time of the exchanges; they do not prove that such records or procedures do not exist internally.
Questions still open
Among the unanswered questions are whether every eligible project received an individual review, whether automated tools helped with early screening, how the judging stages and tie-breakers worked, and whether entrants will eventually receive feedback. The event's Community Vote prompted separate questions about voter eligibility and the precise voting deadline, even though the winner announcement was scheduled for November 16.
The Most Viral award also generated debate over paid promotion on X. Organizers' Discord guidance said paid boosting was allowed while fabricated engagement was not. Some participants argued that allowing advertising favored entrants with larger marketing budgets; others defended their use of ads as compliant with the rules. This disagreement concerns the design of the promotion-based award and, on its own, does not establish manipulation of the results.
One participant responding to a discussion on Reddit said they had received “zero feedback.” That is an individual account, not a representative survey. Taken together, the Discord questions and public comments show a demand for more information from at least some builders, rather than proof of misconduct. The distinction is between having published judging criteria and explaining how those criteria were applied across thousands of entries. A fuller account of the screening stages and whether individual feedback will be available could help resolve those questions while recognizing the achievements of the winning teams.
Reporting note (November 10, 2026): This section is based on participant messages and Cerebral Valley's responses shared in the event's Discord, plus an individual Reddit comment. The public explanations described here may be updated by the organizers.
Sources and editorial methodology
Rankings, amounts and project demonstrations were taken from the November 9 ceremony, including its results slides and recorded presentations. Technical performance figures are attributed to participants and have not been independently audited. Event background: Meta; Cerebral Valley; project gallery. The Community Vote was scheduled for November 16; this report does not invent a winner.
Official project pages for every named winner
Direct links shared by the organizers on Discord, grouped in the gallery.
Comments 0
No comments yet.
Leave a Comment