https://developers.google.com/edge/gallery
Anyone with a 16GB Mac — that is quite a lot of journalists, surely — can download that, install a model into it, and play.
Surely journalists have to start asking questions at least about OpenAI's consumer revenue projections now.
I am a major, major AI cynic, but I decided to be an informed cynic so I've been playing with local models for agentic work and a bit of CAD-to-image generation. I really quite like the 26B Gemma model — I've been using it to teach myself some fundamental things and learn OpenCode without developing a cloud dependency. It writes fairly good code and it is helping me learn the things I want to learn at a pace that I prefer.
But if this 12B model is even half as close as they say it is, this casts some doubt on the consumer end of the cloud business model, at least in the short term.
(Not clear if this app is using the MTP drafters; I've still not got them working with Gemma myself, though the Qwen 3.6 built-in MTP support is super in LM Studio)
However, on my 18GB RAM MacBook Pro, selecting Gemma-4-12B-it results in this error:
> The model "Gemma-4-12B-it' requires more memory (RAM) than is available on your device.
So yeah, my questions about the 16GB marketing copy are fair.
(Though perhaps it'll squeeze in with a small context window? Not sure I understand that aspect yet)
It does seem to use MTP, yes, and it is quite quick — seemingly the underlying LiteRT stuff can do MTP with Gemma 4 and presumably MTP is a big part of the practicality picture here.
The system prompt thing was a surprise when I poked around.
The combination of these things, though, I still think is significant. It’s a product from an old-fashioned (!) FAANG that installs as easily as Chrome, downloads a model as easily as it could be, combines a chat interface with audio and video analysis/transcription, has a customisable system prompt, MTP, agent skills support etc.
Now, it is from Google so they could kill it when they get bored! But clearly this is local AI packaged in a really accessible format, and the model seems quite capable for its size. It is something Microsoft could do when they can really point to easy consumer hardware that can do it well. It’s certainly something Apple could do better with their distillations of Gemini under the Google deal.
I think a sane line of enquiry for a tech journalist is: 1) doesn’t this threaten the appeal of consumer-tier subscriptions to ChatGPT (which is a big part of OpenAI’s revenue plans), and 2) is it therefore not questionable that the buy-and-hold economics of DRAM, SSD and GPU products that OpenAI benefits from having provoked into causing ridiculous price increases is fundamentally anti-consumer?
FAIR did this 2 years ago now: https://arxiv.org/abs/2405.09818
I've been waiting for something like this to be released since then.
The annoying thing is that chameleon was multi-modal out based on the same principles, but this model is just inputs... (I'm curious how they did pre-training without having multi-modal outputs as well. I wonder if they just chopped them off rather than support image output).
Vision embedder (35M parameters): Replaces the 27 vision transformer layers of the other medium-sized Gemma 4 models. Raw 48x48 pixel patches are projected to the LLM hidden dimension with a single matmul. A factorized coordinate lookup (X and Y matrices) attaches spatial location information directly to the input
Standard approach for training MM-LLMs is we train the encoder first, there are O(2-10B) good images on the internet, so encoder needs to see each image O(10-100) times, that is O(100T) tokens, which is more than the entire pre-training budget for most runs. That is the reason we train the encoder separately (smaller model, 2B active vs 30B or 200B active LLM); there is nothing magical about training the encoder and LLM together, it is just more token-efficient to train the image modality first.
After that a s1/s2 system: fast generation, slow wave correction / observation operating over the fast generation seems like the next leap forward.
Tokens create and hide too many problems to be the 'optimal' solution.
Your problem isn't with tokens, but with "language". Tokens have little to do with language, other than usually being consumed in sequence, but that's true of anything that has to span over time. Thinking of tokens as letters or subwords is mistaking the general with the specific. We may have started with letters and words and subwords (trying to find the best balance for training), but then people figured why not add pixel patches to the dictionary, and then sounds, and then other signals, and after iterating on it a bit, we now have image and sound and symbol sequence data all being part of the same token space.
LLMs stopped being about "language" - in the sense of English, or C++ - long, long time ago. We're still using tokens, but they're more like quanta of sensory input now.
You can take it in two directions, I guess - either consider "Large Language Model" to be an anachronym, a name that couldn't keep up with times, but we got used to it back when it made sense, or alternatively, just broaden your understanding of "language" to encompass any pattern of quantized sensory inputs, regardless of modality :).
(Given how we know humans can communicate with pictures, gestures, body language, noises, movement, actions, or even gaze, and that when it becomes common enough, such systems develop their own pattern structure - dare I say vocabulary and grammar - and that none of it requires or usually involves going through a "normal language" intermediary - I'd lean towards the second direction :)).
--
ETA: also wrt. "thinking with tokens", LLMs don't really think in tokens. You may have heard that phrase, that may have been coined by Karpathy, that "for LLMs, tokens are units of thinking". It's a useful shorthand to remind people that prompting models to be terse and skip prose is effectively dumbing them down, but it's also a bit misleading.
A better analogy is that tokens act like clock signals: each consumed token causes certain amount of computation happen in the network, much like a single clock signal in digital electronics, or turning a crank one revolution in a mechanical contraption. This makes tokens "units of thinking" in the sense that processing N tokens causes M amount of computation to happen. Now, for whatever problem you're solving, there is a minimum amount X of computation that is required to solve in correctly, and it's mathematically impossible to do with less. So if you ask an LLM to solve it, it needs to process at least as many tokens as it takes for M = X. If you force the model to be so terse that it makes M < X, you literally make it impossible to succeed. In practice, you need M >> X.
My understanding of pixel representation is: slice a grid in an image, each square slice gets projected into a number array of x long (not sure how long x is, or if it's variable), which then gets projected down to a token representing that space (3-4 long as alpha-numeric) and AGAIN gets passed into "position detector" which outputs a token representing that pixel/position. which gets passed into the lmm (at a significantly reduced/translated signal into token space).
First, before continuing: do I have that mostly correct?
There is no such projection step. The array of x numbers is the token. For text, there is a one-to-one correspondence between the textual representation of a token, its index in the vocabulary of the model, and the array of x numbers that is fed into the linear algebra of the model, so people often equivocate between them; but for images or sound, there is no discrete vocabulary and no textual representation, only the array of x numbers.
12b means 12G @ 8 bits/param (basically lossless) and 6G at 4 b/p (generally accepted 'pretty close' level). Not too bad?
But TBD how well the base model performs before thinking too much about quantization
> Audio: We simplified audio processing even further. We removed the audio encoder entirely and projected the raw audio signal into the same dimensional space as text tokens.
From the visual guide, there's still the 35M parameter embedder, then the linear projector, for vision, and the linear projector for audio, so it does have some parameters used for the multimodal input to project it into the LLM latent space: https://newsletter.maartengrootendorst.com/p/a-visual-guide-...
And the Unsloth quants, which are missing this, don't support multimodal input. (edit: actually, I may have just needed to update my llama.cpp, will check with an updated llama.cpp soon)
I'm downloading the ggml-org GGUFs now, I tried Unsloth but got some weird problems, double checking with the bf16 model to see if the issue was just the quant.
Still progress, but not quite democratic yet.
Weird though that Google might be cannibalising it's own AI subscription service?
https://github.com/ggml-org/llama.cpp/pull/23398
Please don't use Ollama, it's a bad actor in the OSS community.
But I've moved on from Ollama for the time being, though I am mainly interested to see what the Gemma 4 MTP speeds are like on my M1 Max, so I may test it.
I am quite impressed with the tools in LM Studio, which is also a beautiful app, but it is not open source (which challenges my personal strategy somewhat) and I dread its inevitable enshittification.
Nevertheless the GUI has been very helpful while I learn, and I will probably use it until something else presents or my usage pattern settles down from experimentation to something a bit more routine.
I will try oMLX, too, but judging by the LiteRT page I may soon be able to just use that for the larger models if I end up settling with Gemma 4.
You have probably convinced me to give it a try, to be honest.
It's just that, to cut a long story short, I am currently recovering from a level of burnout so severe that twelve months ago had me fully convinced I was actually in early-onset cognitive decline (I am a bit over fifty).
Only a little over two months ago I was still sure I'd have to quit IT and find a slow job because I was so out of the loop; this whole industry shift even in just the last few months is so shocking and strange.
So I have to be a bit cautious about how many indirections I add, if that makes sense. But I am compiling bigger projects than llama.cpp so I will give it a go.
Thank you for the extra detail.
So much to learn but this news has really vindicated my decision to direct my limited span of concentration and focus to learning how to use open weights models and opencode.
That being said, the real value in paid plans is that you get ecosystem integration that can read your gmail, photos, docs, and so on.
Isn't that just projecting the patches into the d_model size vectors that the models takes?
>I am assuming that involves of quantization
12B model in 16GB seems very reasonable to me, int8 is top quality for running models.
12B at int8 would take up 12G memory, or 75% of the system memory which technically fits within 16GB but the OS will not like that. EDIT: On my 18G memory MacBook Pro, LM Studio reports a "partial GPU offload" for the int8 MLX weights. Can't test because the `gemma_unified" architecture is NYI.
The part I hate though is that I’d bet none of the performance claims are based on int8.
Why do we care about bf16 benchmarks when no one will be using that with this model.
It sounds like marketing spin where the performance claims are based on BF16 and the “runs in 16GB” claim is on a totally different quantized version.