← Back to Signal
Technical PositionInference OptimizationPosition on Gemma 4

Composable by design

Why Google's Gemma 4 validates PrizmalRun's thesis, strengthens our market position, and creates a new deployment surface for SpeCo.

Audry Larocque and Jonathan Pilault · April 6, 2026 · 3 min

Key takeaways
  1. 01Gemma 4's efficiencies are real but frozen at training time. They cannot adapt to deployment context or token-level difficulty.
  2. 02PrizmalRun operates at the execution layer, modulating compute per token at runtime. It is orthogonal to Gemma 4's architectural decisions.
  3. 03Stacking PrizmalRun on Gemma 4 compounds gains: structured pruning, dynamic layer skipping, adaptive KV compaction, and precision modulation.
  4. 04Edge variants cannot manage thermals, battery, or per-token precision. PrizmalRun fills this gap with one model on any device.
  5. 05Gemma 4 expands Prizmal's TAM and validates the thesis. It does not touch the differentiation, patent position, or R/E metric.

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.

Model · Gemma 4 weights

Architecture decisions frozen at training time.

Prizmal · PrizmalRun

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.

Gemma 4 · architecture-level efficiency

Fixed sliding and global attention ratio, shared KV cache layers, MoE expert routing, PLE conditioning. Decided before deployment. Cannot change at runtime.

Engine · Qualcomm, Intel, AMD, NVIDIA

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.

DimensionGemma 4PrizmalRun
Acts onModel architecture, fixed at trainingCompute graph, modulated per token at runtime
AttentionFixed 5:1 sliding and global ratioDynamic head reduction and context span per token
KV cacheShared layers, fixed NDynamic compaction, sparse retention per token
SparsityMoE, 3.8B of 25.2B active, fixed routingSpeCo structured pruning: 50% sparsity, 99% reasoning retained
Layer executionAll layers execute for every tokenDynamic layer skipping based on difficulty
PrecisionFixed bfloat16Dynamic FP16 and INT8 modulation per token
Energy managementNot addressedPower-budget scheduling, thermal management
Edge strategySeparate model variants, E2B and E4BOne model, any device. Runtime adapts to hardware.
Quality guardNone, static architectureLogit 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.

Thesis validated

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.

Differentiation intact

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.

TAM expanded

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.

Edge story strengthened

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 structural ceiling Gemma 4 cannot cross

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.

Run the numbers on your own traffic

The R/E Audit measures reasoning per unit of energy on your real workload, read only and reversible.