DARPA's FY26 SBIR Release 4 Just Opened: FALCON, Art of Novel Signals, and Radiation-Hard Memory — All Closing August 19

DARPA does not run its SBIR program the way most of the government does. Where NSF and the Department of War batch topics into a small number of large annual solicitations, DARPA releases topics on a rolling monthly cadence across the fiscal year, each with a compressed roughly four-week open window. The agency's fourth FY2026 release pre-released July 1, opened for submission July 22, and closes August 19. If you run a deep-tech small business and you have been waiting for the "SBIR season" to begin, this is a reminder that at DARPA the season never really stops — and that the firms who win are the ones tracking the pre-release calendar, not the ones discovering a topic three days before it closes.

Release 4 is a good illustration of DARPA's range in a single window. It spans applied artificial intelligence, national-security signals analysis, and hardened semiconductor hardware — three problem areas with almost nothing in common except that DARPA wants a small company to take a swing at each. Here is what the marquee topics actually ask for, and how to think about competing.

FALCON: fusing LLMs with classical statistics

Fusion of Abstract Learning and Context-Optimized Neural-methods (FALCON) is the topic that will draw the most attention, because it sits squarely on the frontier that every technical field is wrestling with right now: how to combine large language models with rigorous, classical statistical analysis. FALCON seeks systems that pair advanced machine-learning methods with LLMs to produce computationally efficient tools capable of interactive statistical analysis of large-scale data.

Read that carefully, because the emphasis is doing a lot of work. DARPA is not asking for a chatbot that talks about statistics. It is asking for a system where an analyst can interrogate a large dataset conversationally while the underlying machinery does defensible, efficient statistical inference — the LLM as an interface and reasoning layer over methods that actually hold up. The hard problems buried in that sentence are the ones a strong proposal will name directly: hallucination control when the model is reporting quantitative results, computational efficiency at scale, and a way to keep the statistical conclusions trustworthy rather than merely fluent.

For a small AI firm, FALCON is attractive precisely because it is not a pure foundation-model play — you are not competing with the labs spending billions on frontier pretraining. You are competing on the integration layer, on efficiency, and on the discipline of making generative systems produce answers an analyst can defend. That is a domain where a focused team with real applied-statistics depth can beat a much larger, more generic competitor.

Art of Novel Signals: forecasting from messy, multilingual data

The second AI-adjacent topic, Art of Novel Signals: Predicting and Forecasting with High Confidence, targets a different technique: Temporal Knowledge Graph Forecasting using in-context learning from novel, multilingual, and multimodal data. In plainer terms, DARPA wants to predict future events by building and reasoning over graphs of entities and relationships as they evolve over time — and to do it with high, quantified confidence from data that is noisy, comes in many languages, and mixes text, imagery, and other modalities.

This is a national-security forecasting problem wearing a machine-learning hat, and it rewards a very specific combination of skills: temporal graph modeling, in-context learning, and the messy reality of multilingual and multimodal ingestion. The phrase "with high confidence" is not decorative — a proposal that can articulate how it will calibrate and communicate uncertainty, not just produce a prediction, is speaking DARPA's language. Forecasting systems that are confidently wrong are worse than useless to an intelligence analyst, and the topic's framing signals that DARPA knows it.

Non-Volatile Memory for Extreme Environments: the hardware bet

The third topic is a reminder that DARPA is not only funding software. Non-Volatile Memory for Extreme Environments seeks a temperature-hard and radiation-tolerant NOR Flash memory system — memory that keeps its data through heat, cold, and radiation that would corrupt or destroy commercial parts. This is the kind of unglamorous, foundational hardware problem that underpins space systems, hardened defense electronics, nuclear instrumentation, and deep-well or high-altitude sensing.

For a semiconductor or advanced-materials small business, this topic is a different competitive calculus than the AI topics. The barrier to entry is higher — you need genuine device physics and fabrication capability, not just a clever algorithm — but the field of credible competitors is correspondingly thinner, and the path to a defense procurement relationship is more direct. Radiation-hardened components are a chronic supply gap, and a company that can demonstrate a manufacturable, qualifiable part is solving a problem the government will keep needing solved for decades.

How DARPA's SBIR mechanics actually work

Understanding the vehicle matters as much as understanding the topics. A few structural facts shape any DARPA SBIR strategy:

• The window is short and fixed. Release 4 opened July 22 and closes August 19 — a roughly four-week runway. DARPA's topics are narrow and technical, and a competitive proposal cannot be written from scratch in that window. Winners start when the topic pre-releases (July 1 here), using the pre-release period to engage the Topic Author with technical questions before official communications lock down at open.
• The pre-release Q&A period is a real advantage. During pre-release you can contact the topic author directly. Once the solicitation opens, that channel typically closes and questions route through a formal, anonymized process. Firms that use pre-release to sharpen their understanding of what the program manager actually wants write measurably better proposals.
• Direct-to-Phase-II can change the whole play. DARPA frequently offers Direct-to-Phase-II (D2P2) options on topics where a company can show it has already completed the equivalent of Phase I feasibility work with prior funding. A D2P2 award skips the small Phase I and moves straight to a substantially larger Phase II — a transformative on-ramp for a firm with relevant prior R&D. Read each topic's instructions closely to see whether D2P2 is available; we covered how decisively this reshapes strategy in our analysis of DARPA's BTO FY26 topics.
• DARPA SBIR is topic-driven, not open-ended. Unlike agencies that fund broad areas of interest, DARPA wants a solution to this specific problem. Proposals that reframe a company's existing product to loosely fit a topic tend to lose to proposals that engage the topic's actual technical requirements head-on.

How a small firm should approach Release 4

If any of these three topics maps to your capabilities, the timeline discipline is straightforward and unforgiving. With the window closing August 19, the proposal should already be well underway — the pre-release period is over, and the productive engagement with topic authors has closed or is closing. The task now is execution: a technically specific work plan, a credible team, a clear articulation of the innovation, and a commercialization story that shows where the technology goes after the SBIR.

The broader strategic point is about cadence. DARPA's monthly release schedule means that missing Release 4 is not a year-long setback — Release 5 topics will pre-release on a similar rhythm, and the smart move for a firm that is not ready this month is to get on the pre-release calendar now so the next relevant topic arrives with weeks of runway instead of days. The single biggest predictor of SBIR success at DARPA is not the quality of your technology in the abstract; it is whether you saw the topic at pre-release and engaged the program before the window opened.

This release also lands in a notably active SBIR year. Congress reauthorized the SBIR and STTR programs through 2031 in April 2026 after a six-month lapse, NSF has restarted its own SBIR machine with a $250 million FY26 allocation and a July 27 deadline, and the Department of War is running its own July-to-August windows. For a deep-tech founder, the non-dilutive capital available across the federal SBIR system right now is unusually deep — and DARPA, with its rolling cadence and its appetite for genuinely hard problems, is the agency that rewards the firms paying closest attention to the calendar.