hunch runs EmbeddingGemma2 as a CPU System One endpoint at 20 ms on text and 100 ms on images

Matt Mastracci's hunch turns EmbeddingGemma2 into a CPU System One endpoint at 20 ms on text and 100 ms on images. Cloudflare Clef teaches a local student in real time. Open-source launch was planned for later the same week. This desk has no public repo or hosted call yet.

Matt Mastracci posted on October 27, 2026, at 15:09 UTC, that he had been cooking on a way to use Google DeepMind’s EmbeddingGemma2 as a real-time /v1/systemone endpoint. The printed clocks are 20 ms on text and 100 ms on images. The post says the setup is trainable in real time, can distill an upstream teacher, and is multi-modal by default. The share image on this page is the thumbnail from that about-30-second video.

Eight minutes later he named the project hunch, said launch was later that week, and said it runs on CPUs rather than GPUs, so a System One endpoint can sit on smaller, cheaper virtual machines without GPUs.

Real-time student and teacher

At 21:47 UTC a second video thread shows live correction. The text says Cloudflare’s Clef helps refine the local model while EmbeddingGemma2 runs on the CPU. Follow-up posts say hunch also decides on text such as log files and agent traces, trains a student from EmbeddingGemma2 embedding vectors, lets a confident student graduate and then check in only occasionally, and calibrates what the teacher knows. A later reply says hunch starts as a general model you can ship and train as needed. Open-source release was planned for the next day in that thread.

EmbeddingGemma 2 underneath

Google’s EmbeddingGemma 2 blog, dated October 6, 2026, names a 740M Apache 2.0 multimodal embedding model on the Gemma 4 architecture. Modular loads cover text at 270M, vision at 170M, and audio at 300M, with a shared 768-d space and an 8K context window. Weights are on google/embeddinggemma-2. Those embedding scores stay on their own board. They are not a Decision Index Full score and not a JevBench composite.

What this desk opened

Opened October 28, 2026 UTC, public search found no hunch GitHub repository and no hosted POST under that name. This desk did not install weights or call a hunch endpoint. The 20 ms and 100 ms figures stay Mastracci’s published claims. His earlier DiffusionGemma-as-Jev line is already on /news/diffusiongemma-as-jev/. Hosted and open Clef stay on /news/cloudflare-clef/ and /news/clef-omni/.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

Primary post is Matt Mastracci, @mmastrac, October 27, 2026, 15:09 UTC, status 2115098609356284223. The text says he has been cooking on a way to use Google DeepMind's EmbeddingGemma2 as a real-time /v1/systemone endpoint at 20 ms on text and 100 ms on images, that it is trainable in real time, can distill an upstream teacher, and is multi-modal by default. The post attaches a video of about 30 seconds. The still on this page is that video's thumbnail.

Eight minutes later, status 2115100634986975298, he names the project hunch, says launch is later that week, and says it runs on CPUs rather than GPUs so a System One endpoint can sit on smaller, cheaper virtual machines without GPUs.

At 21:47 UTC, status 2115198755263881368 opens a thread with an about-83-second video. The text says real-time learning is like teaching a kid, Cloudflare's Clef helps refine the local model, and EmbeddingGemma2 runs live on the CPU. The next posts say hunch also decides on text such as log files and agent traces, that it trains a student live from EmbeddingGemma2 embedding vectors, that a confident student graduates and then checks in only occasionally, that it calibrates what the teacher knows, and that open-source launch was planned for the next day. Status 2115207180840538220, 22:21 UTC, says hunch starts as a general model you can ship and train as needed.

Google's EmbeddingGemma 2 blog, dated October 6, 2026, names a 740M Apache 2.0 multimodal embedding model on the Gemma 4 architecture, with modular text (270M), vision (170M), and audio (300M) encoders, a shared 768-d space, an 8K context window, and Hugging Face card google/embeddinggemma-2. Cloudflare Clef is the October 1 decision model family already on this site at /news/cloudflare-clef/.

Compare

20 ms on text and 100 ms on images are Mastracci's published clocks for hunch on CPU. They are not a desk rerun and not a board median.

Cloudflare's October 1 Clef blog prints median latency 209.3 ms for Clef and 38.8 ms for Clef-flash against a printed Jev cell of 524.1 ms. Those are a different harness and a different day.

taeold/djev-run's README samples=1 latency of 35 to 60 ms is DiffusionGemma-Jev on an RTX PRO 6000, also from Mastracci's line of work, and it is a GPU path. Celeris-1 Decision's printed one-question median is 67 ms. Do not merge those four clocks.

EmbeddingGemma 2's MTEB Code 78.68 and the 740M parameter count are embedding-model figures from Google's October 6 release. They are not a Decision Index Full score and not a JevBench composite.

Clef as teacher is a training role in the hunch posts. Hosted Clef on Workers AI and the open Clef weights stay on /news/cloudflare-clef/ and /news/clef-omni/.

Terms

hunch: Matt Mastracci's October 27, 2026 harness that serves EmbeddingGemma2 as a /v1/systemone endpoint on CPU, with published clocks of 20 ms on text and 100 ms on images, and a student that learns from a teacher such as Cloudflare Clef in real time. Open-source launch was announced for later that week. This desk has no public repo or hosted call yet.
EmbeddingGemma 2: Google DeepMind's October 6, 2026 multimodal embedding model, 740M parameters under Apache 2.0, modular text/vision/audio encoders, 768-d vectors, 8K context. Hugging Face id google/embeddinggemma-2.
student graduation: In Mastracci's October 27 thread, the local student trains on EmbeddingGemma2 embedding vectors until it is confident it matches the teacher, then checks in only occasionally.