mlx gemma4
On April 13th, someone in the LocalLLaMA community flagged what appears to be early MLX compatibility work for Gemma 4, Google's latest release in its lightweight open-weights Gemma family. MLX — Apple's homegrown machine learning framework built specifically for M-series silicon — has quietly become the infrastructure layer that makes local AI practical for the tens of millions of people sitting in front of MacBooks and Mac Studios. The post didn't ignite a firestorm of replies. It didn't need to. In the LocalLLaMA community, these early compatibility notices are the equivalent of a train departure board updating: understated, functional, and meaningful to everyone who knows which train they're catching.
If confirmed, here is what this means. Gemma 4 running natively through MLX would bring Google's latest model architecture into the hands of Mac-native local AI users without the friction of conversion workarounds or performance compromises from non-native runtimes. That matters because MLX isn't just a convenience layer — it routes computation through Apple's unified memory architecture in a way that genuinely changes what's possible on consumer hardware. A capable Gemma 4 variant running efficiently on a MacBook Pro M4 shifts the conversation about which models are worth deploying locally versus remotely. For developers building applications on top of local inference, it expands the menu. For Google, it extends Gemma's reach into an ecosystem that has historically favoured Meta's Llama family simply because the tooling arrived first. The second-order effect is competitive pressure on that tooling ecosystem — every new model that lands cleanly in MLX makes the platform stickier and raises the bar for what "good" local AI infrastructure looks like.
Watch for an official MLX-Community release on Hugging Face, or a pull request in the MLX-LM repository adding Gemma 4 support — either would upgrade this from a community observation to a verified capability.
<!-- SC_OFF --><div class="md"><p>After seeing all of the rave reviews I was excited to try it as a lighter weight replacement for Qwen3-coder-next-fp8 (single RTX Pro 6000, using
<!-- SC_OFF --><div class="md"><p>Beginner here, now I have a good GPU and ollama using docker. Pulled the Gemma4 weights and was able to add it to cursor using ngrok. </p>
<!-- SC_OFF --><div class="md"><p>Hi all I had a quick question while we wait for llama.cpp MTP implementation, have any of y'all tried Gemma4 MTP models on ollama and or
<!-- SC_OFF --><div class="md"><p>I still have to try Gemma4, I have already installed it A4B 27B and I am so curious to see it's potential with my eyes, what I've tried is
<!-- SC_OFF --><div class="md"><p>Hi,</p> <p>has anybody succeeded in running llama.cpp with Gemma 31b dense and Gemma e4b as draft model, and simultaneously inhibit the voice
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1sja4rc/gemma_4_insists_its_not_running_locally/"> <img alt="Gemma 4 insists it’s not running locally"
<!-- SC_OFF --><div class="md"><p>I want to run only text prompts to Gemma4 with llama.cpp, but I dont want to use CLI or server - I want to have it fully embeded inside my code.
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →