The big story here is the encoder-free part, which I still don't fully understand.
> Vision: We replaced Gemma 4’s vision encoder with a lightweight embedding module consisting of a single matrix multiplication, positional embedding and normalizations.
What's Google's business case for releasing open models? Don't get me wrong, I am grateful and appreciative of these releases. I'm trying to understand how it fits into their bigger picture as a for profit company? Are they not helping competitors build on the novel technology they have developed?
Is it simply goodwill and/or marketing? Or am I missing something strategic?
This is a pretty good update. The demo video is a bit funny though - the tester asks to turn the release into bullet points. okay, the model obliges. then the tester says draft an email with this content. BAM! the LLM turns the content from bullets to passages even though it was not asked and it undid the last good thing that it did. i am not sure if it's an email etiquette to not put bullets in the email.
Is this Mac only? Or is that an Ollama issue that it only supports this release of models on Mac? It seems like every tag with the MLX badge is only supported on Mac[0], and that includes all of the tags in this release.
> Novel unified architecture: No multimodal encoders. The vision and audio inputs flow directly into the LLM backbone.
I would be interested in how this actually works. I couldn't find a description of the model architecture (and I did check the links in the Google blog)
I do enjoy the immediate out of touch signaling with the "runs on your 16gb vram laptop" line. Because everyone has a laptop with 16gb vram, or can just pop out and buy a new one, right?
They already provide E2B and E4B that run on (much) smaller devices, including tablets and phones. This fills the gap in the middle. The bigger Gemma 4 models are excellent for their size, but at 8-bit quantization they need about 64GB of VRAM or unified memory. 48GB for 6-bit. Any lower quantization than that, they start to get notably dumber. So, a 12B is interesting for that middle ground.
Surely they must know the current hurdles, but clearly they know that all the relevant people are monitoring the market for the proper hardware to get and 16GB will be an entry point.
Quite a niche release. The MoE outperforms it on score and will likely be faster thanks to lower active weights. So this really only makes sense for specific ram constrained applications that can’t fit a quantized MoE
Quite aside from the architectural changes, I suppose this is the answer to why Google had such a glaring hole in the (pretrained) Gemma4 model lineup between the Gemma4 4b and Gemma4 26b models!
A model that comfortably fits in 16GB of VRAM (allowing room for context) is a welcome upgrade.
Highly recommend just dropping Ollama. You can download binary releases of llama.cpp for every platform and run them trivially in 5 seconds. Ollama serves no purpose other than to take open source work and rebadge as its own, while providing inferior functionality
The result is decent, but it had a few bizzare/trivial syntax errors I had to fix manually: it would do an extra closing bracket or paren a few times, and wanted to separate function definitions with comma. Not sure what that was about, but otherwise the output run just fine.
So, with those qualifiers, I think it's a decent local coding model. It roughly compares with GPT-4.1 (!!), released 14 months ago, on the output: https://senko.net/vibecode-bench/2025/minesweeper-gpt-4.1.ht... (actually I'd call it better, but those syntax errors...)
I ran the quantized version (4-bit GGUF) on my consumer-grade card with 12G of VRAM and got 5t/s for output. Not for interactive use for coding, but fairly capable model.
To me, it's fascinating how much progress we got in over a year. GPT-4.1 was considered an extremely capable coding model. Now we got something with 12B of params performing roughly the same (in this specific benchmark, disclaimers, etc).
Models this small and this capable bode really well for the usefulness of a PC like the RTX Spark that Nvidia/Microsoft announced this week. 128GB of unified memory will likely be more than sufficient for effective local agentic coding, even if SOTA cloud models will still be even better.
Up until this point, I've found the cost/value to unequivocally favor using a cloud subscription, but I would be lying if I didn't worry that one day OpenAI is going to increase the price for my subscription by 5-10x. I rely on these tools enough that if there is a real viable local option, I'm going to take it.
I've heard the assertion that the Gemma 4 models don't do well with lower quantization. I wonder if the "bizzare/trivial" syntax errors would go away at Q8?
88 comments
[ 2.9 ms ] story [ 65.8 ms ] thread> Vision: We replaced Gemma 4’s vision encoder with a lightweight embedding module consisting of a single matrix multiplication, positional embedding and normalizations.
That's technically encoding, just without using a dedicated model for it like SigLIP? The Developer's Guide elaborates, it's still a 35M layer which I am curious is robust enough. https://developers.googleblog.com/gemma-4-12b-the-developer-...
> Small enough to run locally on consumer laptops with 16GB of RAM, it unlocks powerful multimodal and agentic experiences right on your machine.
I am assuming that involves quantization, which due to the quality loss makes that statement somewhat misleading IMO.
Is it simply goodwill and/or marketing? Or am I missing something strategic?
[0] https://ollama.com/library/gemma4/tags
Edit: MLX being Mac-only is independent of the model being MLX (and therefore Mac) only. The latter is what I am asking about.
I would be interested in how this actually works. I couldn't find a description of the model architecture (and I did check the links in the Google blog)
Think almost like unix pipelines, have used it for many workflows.
A model that comfortably fits in 16GB of VRAM (allowing room for context) is a welcome upgrade.
I'm curious how they pre-trained it... I feel like it must have had audio/image output that they chopped off.
I wonder how hard it would be to add it back on.
Wait, *Excluding Chinese language.
This is ... curious.
P.S. Where is gemma 4 124b?
The result is decent, but it had a few bizzare/trivial syntax errors I had to fix manually: it would do an extra closing bracket or paren a few times, and wanted to separate function definitions with comma. Not sure what that was about, but otherwise the output run just fine.
So, with those qualifiers, I think it's a decent local coding model. It roughly compares with GPT-4.1 (!!), released 14 months ago, on the output: https://senko.net/vibecode-bench/2025/minesweeper-gpt-4.1.ht... (actually I'd call it better, but those syntax errors...)
I ran the quantized version (4-bit GGUF) on my consumer-grade card with 12G of VRAM and got 5t/s for output. Not for interactive use for coding, but fairly capable model.
To me, it's fascinating how much progress we got in over a year. GPT-4.1 was considered an extremely capable coding model. Now we got something with 12B of params performing roughly the same (in this specific benchmark, disclaimers, etc).
Lists of various models I tested: https://senko.net/vibecode-bench/
Up until this point, I've found the cost/value to unequivocally favor using a cloud subscription, but I would be lying if I didn't worry that one day OpenAI is going to increase the price for my subscription by 5-10x. I rely on these tools enough that if there is a real viable local option, I'm going to take it.
Can you instruct it to use a lsp?
I had this with Gemini: in the middle of a C++ program it once said RParen instead of using )
It was easy to fix of course, but it makes you question what is going on inside its head.
Qwen 3.6 on the other hand barely uses any memory at all for its KV cache.