Gemma 4 26B-A4B stalls in local coding loops

Gemma 4’s debut as a Mixture of Experts (MoE) model brings frontier reasoning to the desktop, but its reported stalling reveals a mismatch between benchmark logic and production reliability in agentic loops.

• –The 26B MoE architecture (4B active) achieves impressive tokens-per-second on M2/M3 Macs, yet fails to sustain multi-step execution without manual prodding.
• –Users suspect a breakdown in the agent-to-model handshake, possibly due to inconsistent tool-call termination tokens or local inference engine bugs in Ollama.
• –Despite the "lazy" execution, the model's reasoning on complex adapter pattern tasks remains competitive with proprietary models like Claude 3.5 Opus.
• –The hardware efficiency of the A4B variant is a major win for local development, provided the community can stabilize the autonomous execution loop.
• –This friction highlights a growing gap between model intelligence and the reliable tool-calling required for truly "hands-off" agentic coding.