Gemma 4 template breaks tool schemas

// 45d agoPRODUCT UPDATE

Gemma 4 template breaks tool schemas

A user found that Gemma 4's default chat template strips nested JSON Schema constructs from tool definitions, so nullable `$ref`-based parameters lose meaning before the model sees them. A small Jinja patch that preserves schema structure fixed tool calling across the affected setups.

// ANALYSIS

This is a classic "the model isn't the bug, the serialization layer is" failure: if the template flattens schema before the prompt is built, tool calling can look unreliable even when the weights are fine.

–Common MCP and tool schemas rely on `anyOf`, `$ref`, and `$defs`; dropping those breaks more than just one nullable object
–The fact that Qwen3.5 and gpt-oss worked on the same tool points to a template fidelity issue, not a general inference-engine problem
–Preserving `oneOf`, `allOf`, `enum`, `const`, and nullable arrays or objects should make Gemma 4 much less brittle in real agent stacks
–For developers, the takeaway is to inspect rendered prompts and schema expansion, not just model outputs, when tool calls fail

// TAGS

gemma-4llmagentmcpinferenceprompt-engineering

DISCOVERED

45d ago

2026-04-29

PUBLISHED

45d ago

2026-04-29

RELEVANCE

8/ 10

AUTHOR

EntertainmentBroad43

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE2m ago

LMCache is an open-source KV cache management layer that supercharges LLM inference by sharing and reusing KV caches across GPUs, CPUs, and local/remote storage layers.

LMCache optimizes large language model (LLM) inference by extracting the Key-Value (KV) cache from GPU memory and treating it as a persistent, reusable asset rather than temporary, ephemeral data. By storing the KV cache across a tiered storage hierarchy—including CPU RAM, local disks, and remote backends like Redis or S3—LMCache enables prefix reuse across different queries, sessions, and physical machines. This decouples caching from the inference engine itself, offering integrations with popular platforms like vLLM and SGLang to drastically reduce Time-to-First-Token (TTFT) and boost serving throughput.

VIDEO17m ago

Claude Code evolves into proactive autonomous teammate

Maya Nielan from Anthropic's Applied AI team details the transition of Claude Code from a typical command-line utility into a persistent, autonomous teammate. Instead of merely executing commands on demand, the system leverages background routines to autonomously handle developer tasks such as refactoring, updating documentation, running test suites, and managing git flows without requiring active human supervision.

UPDATE34m ago

Grok Build worktrees spark developer praise

A developer on X praises the fast-paced development of Grok Build, an AI coding tool from xAI, noting that the team has shipped numerous updates recently. They specifically recommend that developers explore the new worktrees feature over the weekend to experience the product's latest capabilities.

Gemma 4 template breaks tool schemas