Gemma 4 Stealth Refresh: Flash Attention 4 on Hopper, Multi-Turn Tool Calling Fixed

Google quietly updated the Gemma 4 weights (no version bump): H100/H200 users get 25–70% faster prefill via Flash Attention 4; multi-turn tool calling that stalled mid-conversation now completes; vision OCR on dense documents improves.

If you self-host Gemma 4 via vLLM, pull the latest build — the updates are live in the HuggingFace Hub checkpoints and vLLM's main branch enables the FA4 backend for SM90+ (Hopper) hardware automatically. If you're on A100 or older, the prefill gains don't apply, but the tool-calling and vision fixes do. There is no new model name or version flag — just pull the updated weights.

To confirm you're on the refreshed weights:

# Pull the latest Gemma 4 checkpoint
huggingface-cli download google/gemma-4-9b-it --local-dir ./gemma-4-9b-it

# Start vLLM with FA4 backend (auto-enabled on H100/H200)
vllm serve google/gemma-4-9b-it --enable-prefix-caching

Why it matters: The tool-calling fix makes Gemma 4 meaningfully more useful as a self-hosted agent backbone for teams who need open-weight reliability without API rate limits. The FA4 throughput gain on Hopper hardware is essentially free if you're already on that hardware tier.

Sources: Gemma 4 update — ExplainX, vLLM releases