Google DeepMind released Quantization-Aware Training checkpoints for Gemma 4 today, in both the standard Q4_0 format and a new mobile-optimized format. The concrete memory numbers tell the story: E2B drops from 9.6 GB at BF16 to 3.2 GB at Q4_0 to approximately 1 GB at the new mobile format, with the text-only variant fitting under 1 GB. E4B goes from 15 GB BF16 to 5 GB Q4_0 (mobile format size not disclosed at release). Five sizes ship with QAT checkpoints: E2B, E4B, 12B, 26B A4B (the active-4B Mixture-of-Experts variant), and 31B. The Q4_0 QAT line works with llama.cpp, Ollama, LM Studio, vLLM, and MLX out of the gate; the mobile format targets LiteRT-LM and Transformers.js. HuggingFace collections live at google/gemma-4-qat-q4-0 and google/gemma-4-qat-mobile.
The mobile format is where the engineering substance is, and it is not blanket low-bit quantization. Google's description: targeted 2-bit compression on token-generation layers, with reasoning layers held at higher precision. Static activation scaling factors are pre-calculated during training so the model does not pay the activation-scaling overhead on device, and channel-wise quantization is chosen specifically because it maps cleanly onto mobile-accelerator hardware design. That is a substantively different choice than uniform Q3 or Q2 quantization, which collapses reasoning quality first. Whether the bet survives depends on workload-specific evaluation, and Google has not published Gemma 4 QAT benchmark numbers at this release. The historical reference cited is Gemma 3 QAT cutting the Q4_0 perplexity drop by 54 percent versus PTQ, which is precedent rather than Gemma 4 evidence. Worth noting: QAT does not change the size at a given format. It improves quality at that size. The order-of-magnitude memory drop from 9.6 GB to under 1 GB is from the format, not the training method.
Two ecosystem threads worth pausing on. First, this is the third Gemma 4 release in four days, and read as a sequence the velocity tells you something. Google shipped the base 12B encoder-free multimodal model on June 3, LiteRT-LM with native MTP drafter support and the 2.2x decode claim on June 5 morning, and the QAT mobile format on June 5 evening. That is "model, accelerator, deployer" in less than seventy-two hours, with the deployer move putting the smallest variant inside the memory envelope of a mid-range Android phone. The laptop-local frontier was a real story two days ago; today's announcement quietly extends it to the phone-local frontier for text generation, and that is a different category of deployment surface. Apps that previously needed a cloud round-trip for any meaningful generation now have an in-process option. Second, the targeted-precision pattern is the part worth carrying over even if you never deploy Gemma 4. Most quantization work has been uniform, and the cost has always been that reasoning-heavy benchmarks degrade faster than generation-token benchmarks. Holding reasoning layers at higher precision while squeezing token-generation layers to 2-bit is a surgical move that respects the asymmetric quality cost. Other open-weights families with similar deployment ambitions can apply the same separation.
Monday morning, if you have been wanting to ship LLM features in mobile apps without the cloud round-trip: the under-1-GB E2B variant is now a realistic candidate, run it through your actual workload before betting an architecture decision on it, and pay attention to whether your task is reasoning-heavy (where the 2-bit generation layers still depend on full-precision reasoning to land the right answer) or generation-heavy with thin reasoning (where the format will work well). If you are running Gemma 4 on a laptop via llama.cpp, MLX, or Ollama, switching to the Q4_0 QAT checkpoint is the no-regret upgrade given quality preservation claims; verify on your benchmark before committing. If you are quantizing your own open-weights models, the targeted-precision-by-layer-role pattern is the engineering lesson to study, the assumption that you should quantize everything to the same bit width is now actively being contested. And as always with vendor-published quality claims and no model-specific benchmarks, the burden is on independent evaluation; the llama.cpp and mobile-deployment communities will produce the comparative numbers within the next week or two and that is the validation that matters.