A 14M-Parameter LLM Tends a Fish Tank on an $8 ESP32 as Drone AI Claims a 3x Speedup

BaseRT claims 11.6 to 36.4 fps on a 12 g drone flight controller

On November 13, 2026, Base Compute's Fabian Waschkowski published results from running its BaseRT runtime on the CoPilot Nano, a 42x42 mm, 12 g flight controller from German maker Multirotor with a 2.3 TOPS accelerator. YOLOv5n object detection went from 11.6 to 36.4 frames per second. The company says the model was not retrained and accuracy stayed the same.

The caveat: this is the vendor's own benchmark on its own project, and the post gives no independent test or code. Base Compute says the gains come from three changes. Capture, NPU inference and post-processing now run concurrently. The model graph was optimized for the NPU with bit-exact outputs. Data movement between stages was cut.

The point for small hardware is that pipeline scheduling, not a new model, can be the bigger lever on a fixed NPU. Base Compute says it expects 45 to 55 fps with more tuning and plans to test small VLA models such as SmolVLA. Those are plans, not results.

llama.cpp reaches v1.0 with a native macOS app

The ggml team announced llama.cpp v1.0 in a Hugging Face post, discussed on Hacker News on November 14, 2026. The thread's headline mentions a native macOS app.

A third-party digest of the v1.0.0 release notes, dated November 13, 2026, lists several changes. Deprecated APIs are removed. The NextN/speculative-decoding API is promoted. OCP FP8 KV caches and a host-memory MoE expert cache are added. There are new text and vision architectures, and llama-server gains an audio speech endpoint. I have not verified each line against the release page, and no benchmarks accompany the list.

For edge builders, the host-memory expert cache is the item to watch. It targets running large mixture-of-experts models on machines with little GPU memory. The removed APIs are a breaking change for anyone embedding the library on a board or phone.

pocket-tank: a 14M-parameter LLM keeps a virtual fish tank alive on an ESP32-S3

The pocket-tank repo from GitHub user mediacutlet describes a 14M-parameter language model, distilled from a 26B teacher, that runs a virtual fish tank on an $8 ESP32-S3. The repo was created on August 24, 2026 and was trending on GitHub on November 13.

The repo description does not say how fast the model generates tokens or how much it is quantized, so I cannot say how usable it is. The headline claim is still concrete: a small distilled model doing a narrow job inside a microcontroller's memory, with no cloud call in the description.

Distilling a large teacher into a model this small is what makes a task like this fit on a $8 chip. It is a toy, but a useful one for seeing where tiny LLMs can already hold a task together.

NVIDIA shows native DLA inference from Python on Jetson AGX Orin

In an NVIDIA Technical Blog post dated November 13, 2026, the company walks through running networks on the Deep Learning Accelerator of a Jetson AGX Orin through cuDLA from Python. The stated motive is running multiple neural networks within limited compute and power budgets.

The post is a first-party tutorial. The excerpt I could read gives no latency or power figures, so treat it as tooling news rather than a benchmark. The practical value is that the DLA, a fixed-function engine separate from the GPU, becomes easier to reach for prototyping. That can free the GPU for other work on the same module.

A DIY RV1126B module reportedly beats a Raspberry Pi CM4 at YOLO

CircuitDigest reported on November 14, 2026 on engineer adouglass52's CM4-format module built on the Rockchip RV1126B, which has four Cortex-A53 cores and a 3 TOPS NPU. The module carries 1 GB of LPDDR4X, 8 GB of eMMC and Wi-Fi 6. I could not find the maker's own write-up, so every number below is self-reported and relayed by CircuitDigest.

On the same carrier and cameras, using the median of three runs, YOLOv8n at 640x640 in INT8 ran on the NPU at 25.4 ms per frame (38.9 fps). The CM4, with no NPU, managed 4.0 fps on its CPU. The CM4 stays about 2x faster in a 4-thread sysbench run and keeps HDMI and PCIe, which the module lacks. The maker puts build cost at about $62 at 1,000 units, against a $71.25 CM4101008 list price. The module is not fully pin-compatible: the second camera borrows the DSI0 pins and runs as 2-lane on a stock CM4 carrier.

Most of today's numbers are self-reported or vendor-supplied, so the linked sources are worth reading for what they do and do not measure.