How Satlyt is building local AI for space with Gemma
How Satlyt is building local AI for space with Gemma
Satlyt uses Gemma to analyze telemetry, diagnose faults, and reduce dependence on ground-based processing under tight power, memory, thermal, and connectivity constraints.
Running AI at the edge already demands a tight balance between power and performance, but space introduces environmental constraints rarely encountered in terrestrial settings. For spacecraft operators, transmitting all telemetry and system-log data back to Earth is often impractical because of bandwidth limits and brief orbital communication windows.
Satlyt builds distributed compute infrastructure for space, providing the software layer that enables spacecraft to process, validate, and analyze data at the source. By running lightweight Gemma models onboard, these systems can condense relevant operational context into compact summaries, providing higher-value insights while reducing downlink volume and ground-processing burden.
Balancing local reasoning with spacecraft constraints
Operating in space exposes flight systems to telemetry anomalies, software exceptions, and tight memory constraints. When operational issues occur in orbit, traditional rule-based parsers lack the flexibility to interpret complex stack tracebacks. Rather than downloading volumes of data for ground operators to parse, AI models can generate a compact diagnostic summary and troubleshooting recommendation. This shift to local inference allows the system to convert raw telemetry and code exceptions into actionable information.
Satlyt evaluated multiple small language models, choosing Gemma for its combined strong reasoning capability and highly efficient footprint. In satellite engineering, these tradeoffs are critical: a viable model must fit on constrained hardware, operate without continuous cloud connectivity, produce reliable diagnostic output, and preserve sufficient headroom for the broader software stack.
Our focus is on making onboard AI practical under real spacecraft constraints. With Gemma, we are evaluating how smaller, efficient models can analyze telemetry, system logs, and faults locally while staying within strict power, thermal, and memory limits. That is critical for space systems where every watt, every byte, and every second of downlink matters.
Satlyt has deployed a quantized Gemma 3 1B model onboard a satellite, running via llama.cpp for local analysis of system logs, software errors, and stack traces generated by onboard image-processing workloads.
To evaluate this workflow, Satlyt conducted fault-injection benchmarks where common software errors were introduced into an image-processing pipeline. In two representative scenarios, Gemma reduced the diagnostic payload from 1,319 bytes to 469 bytes (64.4%) and from 1,318 bytes to 464 bytes (64.8%), while generating at 22.71 and 25.48 tokens per second, respectively. The model produced a concise root-cause diagnosis and recommended action from combined telemetry, system-log, and stack-trace context.
Slide 1 of 2
Out-of-memory crash (Payload observation)
Gemma was fed a telemetry snapshot showing 99.6% RAM utilization and a system log indicating an OutOfMemoryError: Failed to allocate 144000000 bytes from the OpenCV preprocessing script when attempting to load a 12000x12000 image.
Hardware sensor failure
Gemma was fed telemetry showing nominal compute resources alongside a kernel warning: tegra-i2c 3180000.i2c: no acknowledge from address 0x36 and an OSError: [Errno 121] Remote I/O error triggered during a scheduled camera payload capture.
Converting these complex tracebacks into concise, human-readable explanations onboard provides two critical benefits: it minimizes downlink data volumes by extracting only high-value diagnostic details, and it shortens the time-to-insight for ground operators, enabling rapid anomaly response.
Optimizing inference for satellite-class hardware
Satlyt is evaluating Gemma 4 E2B as the next phase of its onboard AI work on an NVIDIA Jetson Orin Nano. Testing is focused on characterizing what local reasoning costs in terms of memory, power, thermal impact, and inference speed. For spacecraft designers, these numbers represent the physical constraints of spaceflight:
- Memory footprint: In the current 4-bit Q4_K_M configuration, Gemma 4 E2B reaches approximately 4 GB peak RAM usage on the Jetson Orin Nano. With 8 GB of system RAM available, this leaves remaining memory for other onboard services including sensor acquisition, communications, and telemetry handling.
- Power budget: During active inference, the processor’s total power consumption rises to ~11W, an incremental draw of ~7–8W over the ~4W baseline. This measurement helps quantify the additional power budget required to support local reasoning alongside other spacecraft workloads.
- Thermal management: The workload results in a modest 3°-5°C temperature rise on the processor. In the vacuum of space, where convection is absent and systems rely entirely on passive radiative and conductive heat paths, keeping thermal dissipation low is critical to prevent hardware degradation.
- Inference performance: Ground tests show a generation speed of 19.08 tokens/second. This performance supports near-real-time local diagnostic workflows on the target hardware without requiring continuous cloud connectivity.
Beyond quantization, Satlyt is working to make individual onboard AI workloads significantly smaller through pruning and distillation. The team is removing modality-specific components that are not required for a given workload and developing more specialized models for tasks such as software diagnostics. For some diagnostic workloads, Satlyt is targeting a greater than 90% reduction in model memory footprint, while its broader optimization work targets an approximately 85–90% reduction in the memory envelope required for individual managed onboard AI services. The goal is to allow more specialized AI workloads to operate concurrently within the same constrained compute environment.
Satlyt uses NVIDIA Jetson Nano hardware to evaluate Gemma 3 and Gemma 4 under constrained conditions.
Paving the way for orbital AI data centers
As operators manage larger fleets, the value of local reasoning becomes more important. Satlyt has software packages prepared for up to 50 spacecraft deployments in 2026, creating opportunities to extend these workflows across a broader range of orbital systems and hardware.
Space systems are one of the clearest examples of why local AI matters. You cannot always rely on cloud connectivity, and you cannot afford to downlink every raw log, stack trace, or telemetry stream. Gemma gives us a way to test useful reasoning directly onboard constrained orbital hardware, which is a major step toward more autonomous space operations.
Satlyt is also continuing to optimize Gemma for tighter resource envelopes, including evaluating instruction-tuned models, fine-tuning, layer reduction, and new inference optimizations. The goal is to expand from individual diagnostic workflows toward more autonomous onboard analysis, anomaly detection, summarization, and multi-agent decision support, while spacecraft operators retain command authority.
These efforts support Satlyt’s broader objective of building distributed compute infrastructure in orbit: virtual AI data centers in space.
More from the Gemmaverse
BetterSpeak brings English language practice offline with Gemma 4 E2B
