01 · The release
A strong architecture. A fixed one.
Google DeepMind released Gemma 4 on April 3, 2026: four open-weight models under Apache 2.0. The family includes a 31B dense model ranked third among open models globally, a 26B MoE with 25.2B total and 3.8B active per token via top-8 routing from 128 experts, and two edge variants, E4B and E2B, optimized for mobile and on-device inference.
The architecture introduces several efficiency mechanisms: alternating sliding-window and global attention at a 5:1 ratio, cutting KV cache by roughly 74%; shared KV caches where the final N layers reuse earlier KV tensors; a parallel dense FFN plus MoE in the 26B variant; per-layer embeddings in the edge models; and dual RoPE for long-context stability at 256K tokens.
Every one of these mechanisms addresses a cost or memory bottleneck in transformer inference. Every one of them is frozen at training time. The model ships with these efficiencies baked in. They cannot adapt to deployment context, workload variation, or token-level difficulty after the model is trained.
02 · The stack
Two layers. Two systems.
Architecture decisions frozen at training time.
Reads attention entropy, logit margin variance, and hidden state norms per token. Modulates compute execution in real time: layer skipping, structured pruning, attention scaling, FP16 and INT8 precision, KV compaction.
Fixed sliding and global attention ratio, shared KV cache layers, MoE expert routing, PLE conditioning. Decided before deployment. Cannot change at runtime.
Silicon execution substrate.
A Gemma 4 31B deployment still executes all 60 layers, for every token, at full precision. Its sliding-window attention and GQA reduce memory cost. They do not reduce compute per token based on difficulty. PrizmalRun is the system that makes that decision.
03 · Comparison
What each system does. Where each operates.
| Dimension | Gemma 4 | PrizmalRun |
|---|---|---|
| Acts on | Model architecture, fixed at training | Compute graph, modulated per token at runtime |
| Attention | Fixed 5:1 sliding and global ratio | Dynamic head reduction and context span per token |
| KV cache | Shared layers, fixed N | Dynamic compaction, sparse retention per token |
| Sparsity | MoE, 3.8B of 25.2B active, fixed routing | SpeCo structured pruning: 50% sparsity, 99% reasoning retained |
| Layer execution | All layers execute for every token | Dynamic layer skipping based on difficulty |
| Precision | Fixed bfloat16 | Dynamic FP16 and INT8 modulation per token |
| Energy management | Not addressed | Power-budget scheduling, thermal management |
| Edge strategy | Separate model variants, E2B and E4B | One model, any device. Runtime adapts to hardware. |
| Quality guard | None, static architecture | Logit divergence monitoring, automatic fallback |
04 · Stacked gains
Neither system alone can match the pair.
- On Gemma 4 31B dense: all 60 layers execute for every token, with no built-in sparsity or layer skipping. SpeCo's structured pruning at 50% sparsity halves per-token MLP compute, and adaptive KV compaction addresses the 22GB KV cache footprint reported at 262K context.
- On Gemma 4 26B A4B MoE: the always-on dense FFN running in parallel with the MoE block is a direct pruning target. SpeCo can prune within the dense pathway on easy tokens while preserving expert routing for hard tokens.
- On Gemma 4 E2B and E4B edge: PrizmalRun adds dynamic thermal management, battery optimization, and precision modulation per token. PLE operates at the representation level. SpeCo operates at the execution level. They stack.
05 · Verdicts
Four verdicts. All structural.
Google invested significant research capital to build inference efficiency into Gemma 4 at the architecture level. Each mechanism maps to a capability PrizmalRun already delivers at runtime.
Gemma 4's efficiencies are frozen at training time. PrizmalRun adapts per token, per layer, per device, in real time. This is a different category of system.
Gemma 4 is Apache 2.0 with day-one support across Qualcomm, Intel, AMD, NVIDIA, ARM, and MediaTek. Every Gemma 4 deployment is a potential PrizmalRun deployment.
Shipping E2B and E4B with LiteRT-LM, Android AICore, and MediaPipe SDK confirms edge AI is not speculative. PrizmalRun's energy-aware scheduling fills the gap those models structurally cannot address.
The sliding-window ratio is 5:1 for every token. The shared KV layers are the same N layers for every context. MoE routing activates the same 3.8B parameter budget whether the token is trivial or requires deep reasoning. Architecture-level efficiency treats all tokens as if they require the same resources. How much compute a specific token requires is not an architecture question. It is a control question.