LlamaCon 2025, Space Llama, and the Cerebral Valley Hackathon

On April 29, 2025, Meta hosted its inaugural LlamaCon in San Francisco. It brought together AI researchers, open-source contributors, infrastructure leaders, and developers to explore the next generation of open foundation models.

I was fortunate to receive an invitation thanks to years of activity within the Meta developer ecosystem. For multiple editions of our Valley DevFest conference websites, I implemented multi-provider authentication supporting not only Google, but also Twitter, GitHub, Microsoft, and Facebook (Meta) login. These started as Google Polymer-based web apps (an open-source project spearheaded by Ukrainian GDG chapters under the GDG-X umbrella) and later transitioned to the Angular-based Hoverboard platform, which I heavily customized with internationalization, custom avatar builders, and PWA enhancements. Combined with my work in VR/AR/XR on the Meta Quest platform, this established my standing in Meta’s developer network.

The Keynote Surprise: Space Llama & Booz Allen Hamilton

The conference opened with Meta CEO Mark Zuckerberg taking the stage to unveil the brand-new Llama 4 model family and celebrate the real-world impact of open-source AI.

(Mark Zuckerberg on stage during the LlamaCon 2025 opening keynote.)

To my absolute surprise, the very first highlight featured in the keynote was from my employer, Booz Allen Hamilton!

Meta showcased Space Llama, a pioneering initiative deploying a customized, fine-tuned vision-language Llama model directly into Earth orbit aboard low Earth orbit (LEO) satellites. Operating an AI model in space presents extreme challenges: tight power envelopes, strict thermal limits, cosmic radiation, and severely constrained onboard memory and compute.

• Booz Allen: Large Language Models in Space and Beyond
• Booz Allen & Meta Press Release
• Meta Newsroom: Space Llama Heading into Orbit

(My Booz Allen Hamilton colleague Dan Wald at the Space Llama booth at LlamaCon.)

At the Booz Allen booth, I met up with my colleague Dan Wald. The team had collaborated deeply with Meta engineers to aggressively quantize, prune, and compress the model architecture so it could run on radiation-tolerant edge hardware in space. I also had the chance to speak with the engineer who designed the satellite’s custom Retrieval Augmented Generation (RAG) system, which allows the orbiting payload to query onboard operational manuals and mission telemetry without relying on intermittent Earth ground-station downlinks.

The booth was swamped with interested attendees all morning. Before heading to lunch, I made sure to supply cookies and food to Dan and the team so they could stay energized while answering questions.

Thematic LLM Engineer Lunch: The Craving for Tiny Models

Instead of the general conference lunch, I was invited to a specialized thematic LLM engineer lunch with practitioners working on model deployment and fine-tuning.

Much of the conversation centered around the architectural direction of the Llama 4 family, which embraces Mixture of Experts (MoE):

• Llama 4 Scout: 109 billion total parameters, with 17 billion active parameters per token (A17B).
• Llama 4 Maverick: ~400 billion total parameters, with 17 billion active parameters (A17B).

While MoE is great for inference throughput at datacenter scale, I voiced a strong craving for a true successor in the “tiny” model tier. For edge developers building on mobile devices, smartwatches (such as Full Android Watches), and embedded systems, models like Llama 3.1 8B and smaller sub-10B dense SLMs (Small Language Models) remain the sweet spot because they fit comfortably in local device memory without requiring dozens of gigabytes of VRAM. Several other engineers at the table strongly agreed, noting that edge deployment and local device sovereignty remain high priorities for many production use cases.

Afterparty & Real Llamas

The evening concluded with an afterparty featuring two real, live llamas that greeted attendees outside the venue.

(Real llamas greeting attendees at the LlamaCon 2025 afterparty.)

The interactive photobooths at the party ran generative styling pipelines powered by the newly announced Llama 4 vision capabilities. The model processed photos in real time—though the generative rendering gave my face a humorous resemblance to Stephen Colbert!

(The Llama 4-powered generative AI photobooth result at the afterparty, giving me an amusing Stephen Colbert likeness.)

The Cerebral Valley LlamaCon Hackathon at SHACK15

The momentum continued that weekend on May 3–4, 2025, at the official Cerebral Valley LlamaCon Hackathon, hosted at SHACK15 inside San Francisco’s historic Ferry Building overlooking the bay.

I teamed up with a stellar group:

• Daniel Green: An exceptionally knowledgeable generative AI expert who has also taught internal AI courses at Booz Allen.
• Rahul: A brilliant Silicon Valley engineer and genuine “10x developer” whose velocity, coding speed, and architectural clarity were inspiring to watch.
• Ashley: An engineer originally from Fresno who had previously considered joining me as a co-lead for GDG Fresno before landing a software role in the Bay Area.
• Csaba Toth.

The Project: LlamaLinked

Our project was LlamaLinked (web repository), based primarily on Daniel Green’s concept for an intentional, spam-free professional networking platform.

Rather than drowning users in endless feeds, connection requests, and algorithmic noise, LlamaLinked introduces a “one connection per day” model. Powered by Llama 4 Maverick (17B active parameters served via Groq for near-instant inference) and Gumloop web scrapers, the system analyzes professional backgrounds, interests, and collaborative goals to curate a single, high-compatibility introduction every 24 hours. The experience is designed to be WhatsApp-first and web-accessible, making high-signal professional networking as frictionless as chatting with a colleague.

Synthetic Data Generation

My primary focus during the hackathon was building the Synthetic Data Generator. Leveraging Meta’s synthetic-data-kit library—which I had just learned about during LlamaCon—I built automated pipelines to generate realistic, diverse user personas, simulated interaction histories, and edge-case evaluation datasets. This allowed us to benchmark the matching and summarization prompts (llama-prompt-ops) and establish safety guardrails before deploying the live service.

Here is a quick walkthrough demo of LlamaLinked in action:

Interesting Hackathon Observations: MoE Pruning and Spatial AI

Watching the other hackathon submissions revealed some fascinating technical experiments:

1. MoE Expert Pruning: One finalist team tackled the memory footprint of Llama 4 Scout. Since Scout activates 1 expert out of 16 per token, the team systematically pruned unused expert weights and measured the degradation curve across multilingual benchmarks. Incredibly, they were able to prune at least 4 experts away before detecting meaningful degradation in multilingual performance. This demonstrated that MoE architectures offer promising post-training pruning opportunities for edge and memory-constrained deployments.

2. 3D Stratigraphy with Llama 4: A solo participant—a geologist—attempted using Llama 4 vision to analyze and reconstruct 3D stratigraphic rock layers from geological cross-sections. While it was a creative application, it highlighted a fundamental limitation of current large vision-language models: true 2D and 3D spatial awareness. Language models excel at semantic comprehension and 2D pattern recognition, but understanding physical depth, spatial topology, and 3D volumetric continuity requires dedicated geometric inductive biases. This is why companies like Tesla have invested engineering decades, synthetic simulation environments, and custom Dojo supercomputers to solve spatial perception.

Final Thoughts

LlamaCon 2025 and the Cerebral Valley Hackathon showcased how fast open-source AI is maturing—from running language models aboard satellites in low Earth orbit to running Mixture-of-Experts architectures for intentional social networking. It was an inspiring week reconnecting with colleagues, building alongside talented engineers, and exploring the future of open intelligence.