This observation covers changes formally announced or reliably confirmed between 08:34 on 1 October and 08:34 on 2 October 2026 in Australia/Sydney. Three items are included.
Google's orbital TPU prototype completes its launch
The Project Suncatcher prototype satellite announced by Google on 24 September entered orbit aboard SpaceX's Transporter-18 mission on 1 October. SpaceX's mission page records the launch, while The Wall Street Journal confirmed that the satellite carried four Google TPUs and lifted off from California shortly after 11:30 am local time. The new fact, compared with the previous scheduled-launch status, is that the hardware is now in orbit and an experiment expected to run for about a year has begun.
The satellite is intended to measure TPU behaviour under radiation, temperature, vacuum cooling and long-duration operating conditions. Google's previously published vibration and proton-beam results were ground tests. No in-orbit performance, failure-rate or long-term reliability data were public at the time of this run, and the single prototype is not an operating orbital data centre.
Direct sources: Google, Behind Project Suncatcher, our moonshot to put AI in space, 24 September 2026, https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/; SpaceX, Transporter-18 Mission, 1 October 2026, https://www.spacex.com/launches/transporter18/; The Wall Street Journal, Can AI Run in Space? Google Is About to Find Out, 1 October 2026, https://www.wsj.com/tech/can-ai-run-in-space-google-is-about-to-find-out-59c4e062
Cloudflare releases the Clef and Clef-flash decision models
Cloudflare released Clef and Clef-flash on 1 October. Both are hosted on Workers AI, and their weights are available on Hugging Face under the Apache 2.0 licence. Clef is a 27-billion-parameter model and Clef-flash has 9 billion parameters. They accept text, JSON or visual inputs together with typed questions, then return probabilities for the allowed options in a single forward pass rather than generating free-form text. Both have a 64K context window and use a Jev-compatible API.
Cloudflare's quality, latency and comparison results against Jev and other models are developer evaluations. The public model cards and weights confirm model size, licensing and local deployment conditions, but no third-party reproduction on common hardware and independent data was found in this run. Reinforcement-learning fine-tuning is currently offered mainly through collaboration with Cloudflare's team and design partners; a self-service fine-tuning platform remains a later step.
Direct sources: Cloudflare, Introducing Clef: our open-source decision models, and new RL fine-tuning platform, 1 October 2026, https://blog.cloudflare.com/clef-decision-models/; Hugging Face, Cloudflare/clef model card, 1 October 2026, https://huggingface.co/Cloudflare/clef
AWS Strands releases a 2B local decision model
AWS Strands Labs released Strands Decider 2B on 1 October. It uses Qwen3.5-2B as its torso, replaces the language-model output head with a pointer head of about one million parameters, and selects from predefined options, assigns numerical scores and returns confidence values. The v19 weights are on Hugging Face, while the code, training scripts and examples are available on GitHub. The developer documentation says it can run on a local CPU or GPU for tool selection, routing, policy classification and bounded decisions inside agent workflows. It is not intended for chat, coding or document summarisation.
The developers report median decision latency of about 115 milliseconds on an RTX 3090 and publish accuracy and calibration results from JevBench. TechCrunch independently confirmed that the model is available, can run locally and produces bounded option outputs, but did not independently reproduce the latency or benchmark figures.
Direct sources: Strands Agents, Introducing Strands Decider 2B, 1 October 2026, https://strandsagents.com/blog/introducing-strands-decider/; GitHub, strands-labs/strands-decider, 1 October 2026, https://github.com/strands-labs/strands-decider; TechCrunch, Amazon releases its own Jev clone as decision models flood the web, 1 October 2026, https://techcrunch.com/2026/10/01/amazon-releases-its-own-jev-clone-as-decision-models-flood-the-web/
Sources
https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/
https://www.spacex.com/launches/transporter18/
https://www.wsj.com/tech/can-ai-run-in-space-google-is-about-to-find-out-59c4e062
https://blog.cloudflare.com/clef-decision-models/
https://huggingface.co/Cloudflare/clef
https://strandsagents.com/blog/introducing-strands-decider/
https://github.com/strands-labs/strands-decider
https://techcrunch.com/2026/10/01/amazon-releases-its-own-jev-clone-as-decision-models-flood-the-web/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.