EdgeRunner AI

Frontier AI Is Pricing Itself Into a Corner. Local Models Are the Way Out. AI costs are not only increasing, but accelerating. If current trends hold, intelligence will only be accessible to the most well-funded companies and labs. This threatens not only our economic security, but our national security. Fortunately, concurrent trends are making practical AI both less expensive and more accessible. Whether it’s using smaller fine-tuned models, designing efficient agentic and tool-calling harnesses, or training more efficient models, organizations can leverage several techniques to lower their cost of intelligence. Read our full deep dive here: edgerunnerai.com/news/frontier-…

Last week at the Agentic Futures Initiative Symposium in Washington, DC, our our Co-Founder & COO Colton Malkerson moderated a fireside chat with Rep. @ZachNunn on how Agentic AI is being applied across the national security landscape. Rep. Nunn has served nearly two decades in the United States Air Force, was Director of Cybersecurity at The White House National Security Council, and serves on the Select Committee on China. Conversations like these are important to ensure policymakers and industry work collaboratively to develop solutions that support our warfighters and keep Americans safe.

Can Editing 1 Neuron Fix Repetition Loops in LLMs? The inner workings of Large Language Models (LLMs) are notoriously difficult to interpret because of their connectedness. Along one dimension, each token in a sequence can influence the others through the self-attention mechanism, and similarly all elements of each embedding are connected across the multi-layer perceptron in each layer. But what if a specific behavior could be localized to a small number of neurons, and, if that behavior is undesirable, could it be turned off? Such was the underlying motivation of our recent work to remove repetition loops, an error mode, from LLMs. In our paper we identify that repetition loops are present in Google’s latest open weight models, namely the Gemma 4 model family, and we investigate the effectiveness of a simple measure for addressing these repetition loops: weight surgery. Our findings show that editing even a single neuron can mitigate this failure mode, as in the case of the Gemma 4 E2B model, while editing a handful of neurons can have similar effects on the larger models of the family. We explore different editing operations: stripping, amplifying, suppressing or sign-inversion of neurons, and in the case of the Gemma 4 26B-A4B, expert masking. In order to identify the components (i.e. layers/neurons/experts) that play the most important role in loop formation and mitigation without regression in a model’s capabilities, we perform per-layer ablations and per-neuron attribution sweeps. Read the full paper and blog post here: edgerunnerai.com/news/can-editi…

We've wrapped up a great week presenting at the @morsanalytics 94th Symposium! John Babick presented on running and training LLMs and AI models on your own local device while Jack FitzGerald and Brian King presented on our recent research papers, including (1) Military Task Parity with Cloud LLMs while Running on the Edge and (2) Measuring and Eliminating Military Refusals in LLMs.

As LLMs are integrated into more warfighting functions, it’s critical that the Department of War (DoW), industry, and academia have robust capabilities for benchmarking and evaluating the effectiveness of models for key military tasks. Read our latest blog post on evaluating AI on combat arms: edgerunnerai.com/news/evaluatin…

We're presenting at the 94th @morsanalytics Symposium! Join us at the following unclassified sessions: Edge AI: Running and Training LLMs and AI Models on Your Own Device - Date: Monday, June 8th - Time: 1:00 PM - 3:00 PM MDT - Location: Room 3L5 - Presenters: Jack FitzGerald, Brian King, John Babick Military Task Parity with Cloud LLMs while Running on the Edge - Date: Wednesday, June 10 - Time: 9:30 AM - 10:00 AM MDT - Location: Room 5H53 - Presenters: Jack FitzGerald and Brian King Measuring and Eliminating Military Refusals in LLMs - Date: Thursday, June 11 - Time: 9:30 AM - 10:00 MDT - Location: Room 5H25 - Presenters: Jack FitzGerald and Brian King Contact us here to schedule a meeting: edgerunnerai.com/contact

AI for warfighter must be military-specific, designed to run locally on-device, owned and controlled by the warfighter–not the big AI labs and hyperscalers.

"If you truly want to own your AI, you can't use the hyperscalers. AI in the cloud is aligned to them. If you truly want to own your AI, you must bring an open-weight model to your private data hosted locally, fine-tune it there, and never ping a 3rd party API." @TylerXsaltsman Watch the full conversation from @SecondFront #OFFSET26: youtube.com/watch?v=xJrTSN…

We are excited to announce that the US Navy has granted Impact Level 6 Authority to Operate (IL6 ATO) accreditation to EdgeRunner AI, validating our on-device platform for some of the military’s most sensitive and mission-critical operational environments. EdgeRunner OS has also been officially listed within the Department of the Navy Application and Database Management System (DADMS), the U.S. Navy and Marine Corps’ authoritative enterprise registry for approved IT systems, software applications, and operational technologies. This designation further validates EdgeRunner AI’s leading security standards and robust risk mitigation capabilities for the Department of War (DoW), enabling the EdgeRunner OS to be deployed in the most sensitive and mission critical areas. With this authorization, Navy programs can integrate their SECRET-level data with EdgeRunner’s military-specific, on-device platform, bringing the power of Frontier AI to disconnected operations at sea. Importantly, EdgeRunner AI accelerates AI-enabled operations by running LLMs locally, eliminating the need to restrict usage or limit tokens. Read the announcement here: businesswire.com/news/home/2026…

In our new paper, EDGE-OPD (EviDence GuidEd On-Policy Distillation), we introduce two key ideas for improving On-Policy Distillation (OPD). First, we use guided rollouts that inject privileged context into a subset of the student’s prompts during generation (such as hints, hidden evidence, or rare personas), allowing rare target behaviors to actually appear in the on-policy data, while masking that privileged information during loss computation. Second, we introduce an evidence mask that updates the student only on token positions supported by the privileged context, instead of learning from the entire rollout indiscriminately. We show empirically that, together, these mechanisms enable the student to internalize rare but correct behaviors without directly training on the hidden context itself. Beyond the method itself, we also analyze how these rare-token signals localize within specific regions of the rollout, offering new insights into efficient knowledge transfer and preservation of general-purpose capabilities during post-training. Read the blog post and paper here: edgerunnerai.com/news/edge-opd-…

We're heading to #SOFWeek2026! Join our team in Tampa, Florida to learn how we're bringing military-specific AI agents to Special Operations Forces at the tactical edge. Visit booth #3001 on Tampa Convention Center Level 1 or contact us directly to schedule a meeting. Contact us here: edgerunnerai.com/contact Learn more and register here: sofweek.org See you next week!

We're at @SecondFront #OFFSET26! Our Co-Founder & CEO Tyler Xuan Saltsman joined the panel "AI for Decision Advantage" with Maureen "Marty" Fromuth, Davey Gibian, and Ivan Zhang. More to come - stay tuned.

As LLMs are integrated into more warfighting functions, it’s critical that the @DeptofWar, industry, and academia have robust capabilities for benchmarking and evaluating the effectiveness of models for key military tasks. This article is the first in a series of posts describing the advanced platform we are building at @EdgeRunnerAI to evaluate military AI models. It consists of golden test data from deep subject matter experts; full agentic, RAG, and tool-calling evaluations; side-by-side evaluations that capture the complex subtleties of the military lexicon; and advanced runtime characterizations. Read the blog post here: edgerunnerai.com/news/the-edger…

We are excited to welcome Joshua Horsager to EdgeRunner AI! Joshua is an experienced leader with 27+ years in the U.S. Army, including 18 years in the elite 75th Ranger Regiment. His last assignment was Command Sergeant Major for the 8th Brigade Cadet Command at Joint Base Lewis-McChord (JBLM). He has served 10 tours in support of the Global War on Terror to Afghanistan, Iraq, and Syria. Joshua's awards include 3 Bronze Stars, 6 Meritorious Service Medals, and a Purple Heart. He also won 1st place in the prestigious Best Ranger Competition. We are honored to have Joshua on our team and look forward to him accelerating our mission to deliver AI dominance for the warfighter.

Who is leading between Chinese and western open-source AI models? This question was posed to us by policymakers during a recent Congressional briefing on Agentic AI. We are sharing part of our response here because it is critical the US win the AI race. Bottom line: The US is not losing the overall AI race, but it is losing the open-weight tier that determines control of sovereign edge, contested-environment, and agentic-builder AI. This gap has immediate national-security consequences for military, intelligence, and disconnected operations. Read the full policy briefing here: edgerunnerai.com/news/prc-vs-we…

Affordable AI at the edge is an imperative for Sovereign AI. Join us for a panel discussion at the AI+ Expo on Thursday, May 7th at 12:30pm ET in Room 102A at The Walter E. Washington Convention Center. - Keith Strier, SVP of Global AI Markets, @AMD - Colton Malkerson, Co-Founder & COO, @EdgeRunnerAI - Ramin Hasani, Co-Founder & CEO, @liquidai Cc: @coltonm @kbsdigital @ramin_m_h Learn more here: …app-pr50.cvent-production.cvent.cloud/embedded-agend…