Blog

How Gemma Gemini routes each message, where each model runs, and what we are learning about small models, embeddings and local AI at Offline Base.

  1. Gemma Gemini

    Introducing Gemma Gemini

    A free chat app where a router sends each message to Gemma 4 E2B on a Jetson Orin Nano or to Gemini 4 Argon in the cloud, and shows you why.

  2. Gemma Gemini

    How Gemma Gemini decides who answers each message

    The path of one message through Gemma Gemini: health check, EmbeddingGemma 2 router, two 50% bars, Exa search, failover and the route trace.

  3. Models

    Gemini 4 Argon: what Google released

    Google announced Gemini 4 Argon on September 30, 2026 and made it generally available on October 13. Specs, prices, access and benchmarks, with sources.

  4. Gemma Gemini

    Why Gemini 4 Argon is the big model in Gemma Gemini

    How Gemma Gemini uses Gemini 4 Argon at high reasoning for its hardest messages, why it replaced Gemini 3.8 Flash, and how failover works both ways.

  5. Local AI

    Gemma 4 E2B on a Jetson Orin Nano

    How Gemma Gemini runs Google's Gemma 4 E2B as a 4-bit GGUF under Ollama on an 8 GB Jetson Orin Nano, how fast it is and what it answers.

  6. Embeddings

    What EmbeddingGemma 2 is and what it does

    Google's open multimodal embedding model: its size, 768-number vectors, task prefixes, 8K context, license, and where you can run it.

  7. Embeddings

    We replaced our cloud router with EmbeddingGemma 2

    We built a router from EmbeddingGemma 2 and 164 hand-labeled messages. It scored 93% to Jev's 98% on held-out tests and took over on October 13.

  8. Embeddings

    How embedding models turn text into vectors

    Vectors, cosine similarity, contrastive training, shorter Matryoshka vectors and MTEB, with EmbeddingGemma 2 as the running example.

  9. Embeddings

    Embeddings beyond search: classifying, routing and grouping text

    Embeddings can classify, route, cluster and deduplicate text. A worked example in Python: the logistic-regression heads behind Gemma Gemini's router.

  10. Local AI

    Running EmbeddingGemma 2 next to Gemma on an 8 GB board

    Field notes from adding EmbeddingGemma 2 to a Jetson Orin Nano already serving Gemma: a GPU attempt that took the board offline, then a CPU-only fix.

  11. Embeddings

    Two ways to route: a decision model and an embedding classifier

    Jev reads a rubric in the cloud. EmbeddingGemma 2 embeds each message on one of our Jetson boards, and two classifiers in the app score it. How they compare.

  12. Local AI

    Local models in 2026: what fits on phones, laptops, boards and GPUs

    Gemma 4 memory needs by size, the runtimes and GGUF quantizations that load open models, and what our 8 GB Jetson boards run in October 2026.

  13. Local AI

    Why run AI on your own hardware

    What you gain by running AI on hardware you own, what it costs you, and where a hybrid like Gemma Gemini fits, with numbers from our Jetson boards.

  14. Local AI

    Quantization explained with two real models

    What GGUF names like Q4_0 and Q8_0 mean, what quantization costs, and why our chat model runs at 4 bits and our embedding model at 8.

  15. Offline Base

    Offline Base: local AI for everyone

    Offline Base wants local AI to be easy for everyone. Our five beliefs, the products in development, the Jetson fleet we run now, and how Gemma Gemini fits.

  16. Offline Base

    Why Offline Base built Gemma Gemini

    Gemma Gemini puts Offline Base's small-model bet in public, one routed message at a time, and what building it taught us for Base Drive and Base Router.

  17. Offline Base

    Inside the Offline Base Jetson fleet

    Five Jetson Orin Nano boards, a control plane and an OpenAI-compatible gateway serve Gemma 4 E2B and the router behind Gemma Gemini.

  18. tryingopen

    tryingopen: try any open model, no account

    tryingopen.com puts 24 models, nearly all open-weight, in your browser with no account and nothing saved. What it offers and how its router works.

  19. tryingopen

    From tryingopen to Gemma Gemini: how one router became another

    How Gemma Gemini's first router grew out of tryingopen's, what we changed, and how both apps moved the decision onto our Jetsons a day apart.

  20. Gemma Gemini

    Small models can handle most everyday questions

    What 164 hand-labeled messages, live router scores and timings from our Jetson boards show about when a small model is enough.