Open Source AI Summit 2026 Guide: Dates, Speakers, Agenda, and Local AI

The most important open-source AI question in 2026 is no longer whether a laptop can produce a few tokens from a downloaded model. It is whether you can inspect the system, choose where your data goes, replace one component without rebuilding everything, and keep the useful parts running when a vendor changes its price, policy, or API.

That is what makes the Open Source AI Summit unusually relevant. The two-day San Francisco event is not organized around a single model launch. Its agenda moves between local inference, agent frameworks, private computing, verifiable hardware, open science, economics, safety, and the right to compute. In other words, it asks what it would take for people to build, run, and own more of their AI stack.

The Open Source AI Summit 2026 takes place September 10–11 at the Presidio in San Francisco. Applications are now closed and the event has reached capacity. No public livestream has been announced on the event website as of September 10. This guide separates those confirmed facts from assumptions, highlights the sessions most useful to local AI builders, and turns the summit's broad themes into practical decisions for a home lab or private AI server.

Open Source AI Summit 2026 Dates, Location, and Event Status

The summit runs for two days, Thursday, September 10 and Friday, September 11, 2026, at the Presidio in San Francisco. The published program begins with breakfast at 9:00 a.m. on both days. Thursday concludes with a happy hour, while Friday's scheduled talks close at 3:40 p.m., followed by an optional gathering.

Detail	Confirmed information	Planning note
Event	Open Source AI Summit 2026	A two-day, in-person gathering focused on open, local, private, and decentralized AI.
Dates	September 10–11, 2026	September 11 is Day Two, not the opening day.
Location	The Presidio, San Francisco	Accepted attendees should follow their confirmation email for access instructions.
Time zone	San Francisco local time	In September, San Francisco observes Pacific Daylight Time.
Attendance	Application required; capacity reached	The application interface currently states that applications are closed.
Ticket price	$150 for accepted applicants	This was the reservation price, not an open walk-in ticket.
Livestream	Not announced	Do not assume that a remote fireside chat means the whole summit is streamed.

Schedules and event access can change at short notice. Check the official summit schedule before traveling or relying on a particular session.

What Is the Open Source AI Summit?

The Open Source AI Summit is a curated gathering about AI systems that people and organizations can understand, operate, modify, and control more directly. The 2026 program spans 12 connected themes: local AI for business, open-model use cases, inference and compute, agentic frameworks, economics and markets, AI for science, safety and guardrails, verifiable hardware, private AI, the right to compute, open versus closed AI, and the global AI race.

That mix matters. An open model does not become a useful private assistant merely because its weights can be downloaded. It still needs an inference engine, storage, permissions, an interface, context retrieval, tool access, monitoring, updates, and a security boundary. The summit treats those surrounding layers as part of the AI ownership question instead of discussing model benchmarks in isolation.

The event is organized by Presidio Bitcoin, a San Francisco hub focused on private and decentralized technology. That background explains why the program extends beyond engineering into economics, civil liberties, policy, hardware trust, and compute access. Attendees should expect a point of view, not a neutral trade show covering every part of the AI market equally.

Who Is Speaking at Open Source AI Summit 2026?

The event website lists 47 speakers across models and inference, agents and tooling, privacy and security, open science, policy and rights, and philosophy and economics. Notable participants include NVIDIA Vice President of Applied Deep Learning Research Bryan Catanzaro; Hugging Face VP of Product Jeff Boudier; NumPy and SciPy creator Travis Oliphant; Google DeepMind Staff Research Engineer Ravin Kumar; AI2 researchers; and practitioners from Together AI, Block, Spiral, Google, Prime Intellect, OpenTeams, and privacy-focused AI companies.

A long speaker list is not automatically a useful agenda. For local AI builders, the better filter is the system layer each session helps clarify:

If you are deciding...	Prioritize these topics	What to listen for
Which models can run locally	Local AI, inference engines, open-model customization	Memory use, quantization, throughput, context limits, hardware support, and operational tradeoffs.
How to build a useful agent	Agent loops, harnesses, protocols, and tooling	State, retries, tool permissions, observability, model switching, and failure recovery.
How to protect sensitive prompts	Private AI, confidential computing, and security	Where plaintext exists, who controls keys, what can be verified, and which claims require trust.
Whether a model is genuinely open	Open weights, open science, licenses, and model customization	Access to training information, code, parameters, modification rights, and redistribution terms.
Who controls AI infrastructure	Right to compute, economics, geopolitics, and verifiable hardware	Availability, concentration, regulation, hardware trust, and dependence on centralized providers.

Open Source AI Summit Agenda: What to Watch on September 10

Day One moves from the political case for open AI into the mechanics of inference, model customization, local agents, open protocols, and software security. Three sessions create a particularly useful technical sequence.

State of Local AI

Ahmad M. Osman's “State of Local AI” is the most direct entry point for readers interested in running AI on their own hardware. The useful questions are practical: Which workloads have moved from demonstrations into reliable daily use? Where do memory bandwidth and model size still create friction? When does local inference reduce cost, and when does it merely move cost into hardware, power, and maintenance?

How Inference Engines Actually Work

Zain Hasan's inference-engine session should help separate model capability from runtime performance. Two people can download the same model and get very different results because model format, quantization, batching, context length, cache management, backend support, and memory movement all affect speed and stability. “Can this machine run the model?” is therefore incomplete. The better question is whether it can run the required context and workload at an acceptable response time.

Engineering Agent Loops and Graphs with Local Models

Dustin Dannenhauer's session moves from a single prompt to an agent system. This distinction is important. A local model may be adequate for classification, extraction, routing, or a bounded tool call even when it is weaker than a frontier cloud model at open-ended reasoning. Good agent design assigns each model a job it can perform reliably and adds limits around tools, credentials, retries, and state.

The later “Agentic Harness Landscape” and “Securing Open Source Software” sessions add two missing layers: orchestration and supply-chain risk. Downloadable code is inspectable, but it is not automatically safe. Models, containers, extensions, agent skills, Python packages, and community templates all expand the trusted computing base.

Open Source AI Summit Agenda: What Happens on September 11?

Friday, September 11 is Day Two. It has a tighter progression from building with open models to privacy, enterprise sovereignty, science, geopolitics, verifiable hardware, and NVIDIA's view of open AI infrastructure.

Time	Session	Why it matters
9:30–9:55 a.m.	How to Build Your Own AI with Open Models	Hugging Face's Jeff Boudier connects model access with the components needed to turn a model into a usable system.
10:05–10:35 a.m.	Computing Manifesto Fireside Chat	Frames access to compute and user control as social and economic questions, not only infrastructure choices.
10:45–11:15 a.m.	Privacy Panel	Useful for comparing local execution, trusted execution environments, encrypted services, and the claims each architecture can support.
11:30 a.m.–12:00 p.m.	Sovereign by Design	Asks how ownership of infrastructure, data, and software affects enterprise value and long-term dependence.
12:10–12:40 p.m.	AI for Science	Explores how open research infrastructure can make models, methods, and results easier to inspect and extend.
1:45–2:15 p.m.	The Geopolitics of AI	Places chips, models, regulation, and compute capacity inside the broader competition for technological control.
2:25–2:55 p.m.	Hardware You Can Verify	Pushes the trust question below software into firmware, hardware, attestation, and the physical machine.
3:05–3:35 p.m.	Fireside with Bryan Catanzaro	Offers NVIDIA's perspective on model research, inference, and the hardware-software stack supporting open AI.

The day closes at 3:40 p.m. after a short closing segment. An optional happy hour and gathering follows at 38 Keyes Avenue. Attendees should still use their event confirmation for final access details rather than treating an online summary as an admission document.

Open Source AI Is Not the Same as Open Weights

The phrase “open-source model” is often used loosely. A provider may release weights while withholding important training data information, training code, or rights needed to modify and redistribute the system. That can still be highly useful, but it is not the same level of openness.

The Open Source AI Definition focuses on four freedoms: using the system for any purpose, studying how it works, modifying it, and sharing it with or without changes. It also says that the preferred form for modification must include sufficiently detailed data information, the relevant code, and model parameters.

Term	What it usually tells you	What it does not guarantee
Open weights	You can obtain learned model parameters under stated terms.	Full training transparency, unrestricted use, reproducibility, or an OSI-aligned license.
Open-source AI	The system is intended to provide meaningful rights to use, study, modify, and share.	That a particular download is small, safe, unbiased, accurate, or easy to operate.
Local AI	Some or all inference runs on hardware you control nearby.	Open licensing, transparent training, good security configuration, or zero external connections.
Private AI	The architecture is designed to limit unauthorized access to prompts, files, outputs, or identity.	That everything runs locally or that every privacy claim is independently verifiable.

These labels describe different dimensions. A closed model can be delivered through a privacy-preserving service. An open-weight model can run locally. A genuinely open system can still leak data through a badly configured interface or agent tool. Buyers and builders should ask separate questions about licensing, execution location, data flow, security, and operational control.

Why Local Inference Is Only One Layer of AI Ownership

Downloading a model feels like ownership because the most visible dependency disappears. The less visible dependencies remain. A useful AI service needs data storage, a runtime, a user interface, identity, retrieval, logs, backups, network controls, and sometimes access to external tools.

For a home lab, it helps to divide the stack into four layers:

1. Data layer: documents, photos, media, embeddings, databases, model files, and backups.
2. Compute layer: CPU, GPU, memory, inference engine, model format, and acceleration libraries.
3. Service layer: APIs, local AI web interfaces, retrieval pipelines, agent runtimes, and automation.
4. Control layer: user accounts, permissions, secrets, network exposure, logs, updates, and recovery.

Those layers do not have to live in one box. A GPU workstation can perform inference while an always-on home server stores documents, runs the web interface, controls user access, and backs up configuration. This split often makes more sense than forcing a storage-oriented NAS to become a high-end AI workstation.

If you are designing that architecture from scratch, start with a local AI server build and define the workload before buying hardware. A small model serving one person has very different requirements from a multimodal agent indexing a shared media library.

What Should Run Locally and What Can Stay in the Cloud?

The summit's open-versus-closed framing does not require an all-local answer. A better design places each task where its privacy, latency, capability, and maintenance requirements fit.

Workload	Usually favors local execution	Usually favors a cloud model
Private document search	Sensitive source files, repeated queries, predictable scope, and LAN access.	Very large context or advanced reasoning that a local model cannot handle reliably.
Classification and extraction	High-volume, repetitive, bounded tasks with a testable output format.	Rare inputs requiring broader knowledge or stronger multimodal understanding.
Coding assistant	Private repositories, offline work, code completion, and narrow repository questions.	Complex debugging or architecture tasks where stronger reasoning creates more value than local control.
Agent tool use	Local files, home services, or internal APIs protected by strict permissions.	Open-ended research or demanding planning, provided sensitive context is removed or controlled.
Image and media analysis	Personal photo libraries, security footage, private recordings, and background indexing.	Occasional high-end generation or analysis that would require expensive local accelerators.

A hybrid system can route routine private work to a local model and escalate selected tasks to a cloud model only after the user approves the data being sent. The hard part is not connecting two APIs. It is making routing visible, preventing silent data leakage, preserving useful logs without storing secrets, and failing safely when either service is unavailable.

How a Home Server Fits the Open AI Stack

A home server is most valuable when treated as persistent infrastructure rather than a benchmark machine. It can remain available while laptops sleep, centralize model and document storage, host a web interface for multiple devices, run scheduled indexing, and keep backups separate from experimental AI environments.

A low-power system can handle orchestration, retrieval, light inference, databases, and interfaces even when demanding generation runs elsewhere. The low-power AI workload guide explains where CPU-only and modest hardware remain useful. For interface selection, compare local AI web UIs by authentication, model compatibility, retrieval support, and multi-user controls rather than appearance alone.

Storage-heavy builders may use a system such as the ZimaCube 2 personal cloud NAS as the data and service layer, then connect separate GPU compute when larger models require it. That division keeps valuable files and always-on services stable while the faster-changing model environment can be upgraded, rebuilt, or isolated independently.

Five Questions to Take Away from the Summit

Whether you attend in San Francisco or follow the projects and speakers afterward, these five questions are more durable than any single announcement:

1. What do I actually own? Check the model license, source availability, data, configuration, prompts, embeddings, and export path.
2. Where does plaintext exist? Trace prompts, retrieved documents, tool outputs, logs, temporary files, backups, and telemetry.
3. Which component can I replace? Test whether the model, inference engine, interface, vector store, and agent framework are truly interchangeable.
4. What happens when the agent is wrong? Limit permissions, require confirmation for consequential actions, preserve recovery points, and make failures visible.
5. What workload is worth running locally? Measure privacy value, frequency, latency, hardware cost, power, maintenance, and the quality gap against hosted alternatives.

Open AI becomes meaningful when those questions have concrete answers. A downloaded model is a start. A system that you can understand, move, repair, secure, and continue operating is much closer to ownership.

Open Source AI Summit 2026 FAQ

When is the Open Source AI Summit 2026?

The Open Source AI Summit takes place Thursday, September 10 and Friday, September 11, 2026. Breakfast is scheduled from 9:00 a.m. on both days. September 11 is the second day of the summit.

Where is the Open Source AI Summit held?

The event is held at the Presidio in San Francisco, California. Accepted attendees should use the location and access instructions supplied by the organizer.

Can I still get Open Source AI Summit tickets?

No general ticket sale is currently available. Attendance required an application, and the event's application interface states that the summit has reached capacity. The published reservation price for accepted applicants was $150, with a limited number of complimentary passes previously offered to open-source AI builders and academics.

Is there an Open Source AI Summit livestream?

The organizer has not announced a public livestream on the event website as of September 10, 2026. One agenda item is labeled as a remote fireside chat, but that describes the speaker's participation and should not be interpreted as confirmation that the full event is streamed.

Who are the main Open Source AI Summit speakers?

The published lineup includes Bryan Catanzaro, Jeff Boudier, Travis Oliphant, Ravin Kumar, David Sacks, researchers from the Allen Institute for AI, and practitioners from Together AI, Google, Block, Spiral, Prime Intellect, OpenTeams, and other open and private AI organizations.

What is the difference between open-source AI and open-weight AI?

Open weights means that a model's learned parameters are available under stated terms. Open-source AI is a broader claim involving meaningful freedom to use, study, modify, and share the system, together with access to the preferred form needed for modification. A weight release may be useful without satisfying every part of that definition.

Does local AI automatically protect privacy?

No. Local inference can reduce the need to send prompts and files to an external model provider, but privacy also depends on network exposure, authentication, logs, extensions, agent permissions, telemetry, backups, and the security of every connected tool.

Do I need a GPU to run open models locally?

Not always. Small quantized models, embedding models, classifiers, and light assistants can run on CPUs, although response speed and context capacity may be limited. Larger models, multimodal workloads, long contexts, and concurrent users generally benefit from more memory bandwidth and GPU acceleration.

Can a NAS run local AI?

Yes, within realistic limits. A NAS or home server can run smaller models, retrieval services, databases, interfaces, and automation. For demanding inference, it can instead store the data and host the service layer while a separate GPU machine handles model execution.

Why does verifiable hardware matter for private AI?

Software openness cannot prove everything about the machine beneath it. Firmware, boot integrity, device identity, secure enclaves, attestation, and hardware supply chains affect whether a system's security claims can be verified. Verifiable hardware extends the trust question from model code down to physical infrastructure.