Qwen
Alibaba Qwen
One family spans tiny dense models, coding MoEs and conditional memory.
Gemma / Lab profile
Google DeepMind
Gemma is a family of different design choices, not just different sizes.
First time comparing models? Learn the terms before the architecture →
Gemma 4 is a particularly useful learning family because its branches solve different problems. The April release includes effective-parameter E models, a conventional dense model and an MoE. June adds a unified encoder-free multimodal model and an experimental diffusion generator.
Architecture trajectory
What connects these models?
E2B and E4B use per-layer embedding tables, so effective parameters differ from total stored parameters. The 26B-A4B and 31B pair exposes the MoE-versus-dense tradeoff. The 12B model changes the multimodal input path; DiffusionGemma changes the generation loop itself.
Output emerges through repeated refinement.
Measure final-answer quality and completion time.
Autoregressive TPOT is not a universal comparison.
Newest first / grouped sequence, not a time scale
The Gemma timeline
diffusiongemma-26B-A4B-it
A different generation loop, not a drop-in speed setting
The June research release generates through iterative refinement rather than the standard one-token-at-a-time loop. That makes it historically useful even if your current service remains autoregressive.
gemma-4-12B-it
One model path for text, vision and audio
The June addition introduced a unified input path within the Gemma 4 family. It is an architectural branch, not simply a halfway size between E4B and 26B-A4B.
gemma-4-26B-A4B-it
Sparse compute in a workstation-sized family
The April family offered both dense and MoE choices. The reported 25.2B total and 3.8B active values make the rounded 26B-A4B name precise enough to reason about.
gemma-4-31B-it
The dense control for Gemma’s MoE experiment
Its role in the April release is complementary to the MoE. The model accepts text, images and video; audio support belongs to other variants and must not be inferred from the Gemma brand.
gemma-4-E2B-it
“Effective” is a compute label, not the download size
Per-layer embedding tables are the reason the E label differs from a conventional dense parameter label. Placement and quantization of those tables can change the practical footprint substantially.
gemma-4-E4B-it
More effective capacity, with tables still to place
Like E2B, it makes conditional embedding storage visible in a small-model setting. This is a useful bridge to the much larger conditional-memory systems introduced later in 2026.
Which other labs should you compare?
Use Qwen’s small dense models as alternatives to the E variants, while explicitly accounting for embedding storage. Compare Gemma 31B with Qwen3.8-27B or Muse Glimmer for mid-sized multimodal applications. DiffusionGemma needs its own quality-versus-completion-time experiment against autoregressive Gemma.
Muse Glimmer
Meta
A local-agent release includes the runtime around the model.
Interview preparation
What you should be able to explain
Explain three independent choices: where inputs become model tokens, which parameters execute per token, and how output tokens are generated. Gemma offers a concrete example for each.
Study the underlying systems → · Get the illustrated report →
Sources and scope
Checked 13 September 2026. Assessments and comparison candidates are our interpretation of the cited designs, not measured cross-lab rankings. Each timeline entry links to its release evidence; model pages preserve pinned card and configuration sources.