REDDIT · REDDIT// 3h agoMODEL RELEASE

Gemma 4 26B fails local function calls

Google's recently launched Gemma 4 26B Mixture-of-Experts (MoE) model is facing widespread reports of broken function calling and tool execution in local environments. Users transitioning from cloud-hosted versions to local inference via Ollama and llama.cpp report that the model frequently ignores tool-use commands and fails to maintain agentic reasoning loops despite its high benchmark rankings.

// ANALYSIS

Gemma 4's impressive 26B MoE architecture is a speed demon that currently lacks the local stability to handle complex agentic workflows.

–The 26B variant activates only 3.8B parameters per token, which likely contributes to reasoning brittle-ness when parsing complex tool schemas compared to the 31B dense variant.
–Critical bugs in early Ollama v0.20.x releases caused tool-call responses to be misrouted to reasoning fields, a major hurdle for initial local adoption.
–Formatting hallucinations, such as incorrect parameter names and abrupt generation ends, suggest that chat templates and quantizations for the new model are still maturing.
–While Gemma 4 scores 86.4% on function-calling benchmarks, the "last mile" of local implementation remains a significant friction point for developers building autonomous agents.

// TAGS

gemma-4llmai-codingagentmcpopen-sourceopen-weights

DISCOVERED

3h ago

2026-04-20

PUBLISHED

7h ago

2026-04-20

RELEVANCE

9/ 10

AUTHOR

bishwasbhn