Gemma 4: five sizes, Apache 2.0, and the small end matters most
A family rather than a model: from 2.3B effective parameters up to 31B dense, all multimodal, all reasoning-capable, all under a licence with no conditions attached.
The architecture
how it is builtFive sizes, two shapes
E2B at 2.3B effective and E4B at around 4.5B for edge, a 12B unified multimodal, a 26B mixture-of-experts with roughly 4B active, and a 31B dense. The MoE variant is the interesting one for serving: 26B resident, 4B working.
Multimodal across the whole family
Native image and video on every size, with audio input on the two smallest. That is unusual: multimodality normally arrives at the top of a family and trickles down.
256K context
Shorter than the million-token models elsewhere on this site, and I do not think that is a weakness so much as a different target. Doubling the previous generation matters more for the workloads these sizes serve.
Apache 2.0
No thresholds, no attribution requirement, no bespoke terms. On a page where two of the six entries have licences that need reading twice, this is worth stating plainly.
Serving it
the playbookThe whole family is small enough to be uninteresting from a placement perspective, which is the point. The 31B dense at BF16 is around 62 GB and fits on a single GPU; the edge variants run on a laptop or a phone.
On a 4,000-GPU HGX B200 fleet
500 nodes, eight B200s each, 180 GB per GPU
Weights are 62 GB (31B dense at BF16, derived). Usable memory is 180 GB x 0.92 per GPU,
the gpu_memory_utilization from the published recipe. A placement only counts
here if the cache gets at least 30% of what is left, because a config that fits the weights
and nothing else cannot serve a request.
Tensor parallel stays inside the NVSwitch, so no TP traffic ever crosses a NIC.
- The MoE is a drafter candidate. 26B total with roughly 4B active is the right shape for speculative decoding against a much larger target, and Apache 2.0 means you can fine-tune it on your own traffic and ship the result. Given that a drafter trained on your distribution is the one advantage a vendor cannot sell you, that combination is more interesting than the benchmark scores.
- Effective parameters is not total parameters. The E-series figures are quoted as effective rather than total, reflecting per-layer embedding techniques. Do not size memory from those numbers without checking what the checkpoint actually weighs.
- It is the counterexample on this page. Every other entry is a question about whether it fits on a node. This one is a question about whether you need a node at all, and for a large class of production work the answer is no.
The one that changes who can serve anything at all
my readI spend most of my time thinking about models that need a node, so a family whose smallest member runs on a phone is easy for me to dismiss and wrong to. The E2B and E4B variants are the only models on this site that a product team can ship without asking me for GPUs, and Apache 2.0 means nobody has to read a licence first. What I would actually use the 26B MoE for is the drafter role in speculative decoding: four billion active parameters is cheap enough to run beside a large target, and a permissively licensed drafter you can fine-tune on your own traffic is exactly the asset the field says is the real moat.
The numbers
measured elsewhere, not here- Sizes
- 5
- Smallest
- 2.3B eff.
- Largest
- 31B dense
- Licence
- Apache 2.0
Sizes and capabilities from the model card and third-party analysis. Intelligence-index positions come from Artificial Analysis, not from my own evaluation.
Sources: Gemma 4 model card · Artificial Analysis: Gemma 4