Gemma 4 Same-Name Update: FA4 Prefill, Tool Fixes, Re-pull Guide

In July 2026, multiple independent outlets reported that Google shipped a same-name weight and chat-template refresh for the open Gemma 4 family—not a “Gemma 4.1” version bump.

📑Table of Contents

1. What the same-name update actually changed
2. Which GPUs get the FA4 speed claim—and which do not
3. Tool-calling and truncated-response fixes
4. How to re-pull weights—and when you must
5. How to read the agentic benchmark tables
6. Official function-calling verification steps
7. FAQ
8. Author Perspective
9. Summary and next steps

What to verify before trusting the headline

Headlines highlight “up to 70% faster” processing and better tool calling. In practice, three questions decide whether the refresh matters for you:

• Whether your GPU path can use Flash Attention 4 (FA4) on Hopper
• Whether agent tool-call reliability and truncated replies improve on your stack
• Whether you must re-pull Hugging Face weights yourself

This article cross-checks THE DECODER, TechTimes, explainx, AlphaSignal, Google’s official function-calling docs, and the public Hugging Face distribution surface.

Reported numbers are vendor/media claim ranges, not guarantees measured on your hardware.

---

What the same-name update actually changed

This is a silent refresh of existing Gemma 4 weights and templates, not a new major version name.

Three practical deltas

The practical deltas cluster into three areas:

1. Conditional speed: On NVIDIA Hopper (H100-class) with FA4 enabled, reports cite about 25–70% faster prefill and up to about 31% lower TTFT, with the upper end load-dependent.
2. Agent quality: Fixes for tool-calling failures, fewer mid-response truncations (“laziness”), and chat-template corrections around items such as tool_call_id, StrictUndefined, and Jinja templates.
3. Vision headroom: Default max_soft_tokens remains 280; analyses describe optional expansion up to 1120 (roughly ~2.51MP OCR-class resolution) when needed.

---

Product context

Product context still matches Google’s April 2026 Keyword post.

Gemma 4 is an Apache 2.0 open-model family sized across E2B / E4B / 12B / 26B-A4B / 31B (and peers on the HF collection), with agentic function calling as a first-class positioning.

---

Sources

• THE DECODER (2026-07-16)
• TechTimes (2026-07-16)
• explainx
• AlphaSignal
• Google Keyword (2026-04-02)

---

Which GPUs get the FA4 speed claim—and which do not

The “70% faster” line is not a universal GPU claim.

Coverage repeatedly ties the headline prefill gains to Hopper + FA4. Ada-class cards (for example RTX 4090) should not be expected to inherit that headline speedup.

Pure cloud-API users may not re-pull weights at all, but they still need to confirm whether the provider rolled the new revision.

GPU path comparison

GPU / path	Expected effect	Check first
Hopper + FA4 enabled	Prefill ~+25–70%, TTFT up to ~−31% (claim range)	Runtime truly uses FA4 / flash_attn v4-class kernels
Ada / non-Hopper	Do not expect headline speed	Evaluate tool-call and truncation quality only
Cloud API only	Local re-pull often unnecessary	Confirm provider model revision

---

Sources

• TechTimes (Hopper-limited framing; Ada misses the main speed claim)
• THE DECODER
• AlphaSignal (as of July 2026)

For local-model selection that starts from VRAM and measurement rather than marketing, see our note on practical local LLM choice.

---

Tool-calling and truncated-response fixes

If Gemma 4 sits on an agent path, tool reliability and incomplete replies usually matter more than prefill marketing.

Independent reports list chat-template fixes (tool_call_id, StrictUndefined, O(n) backward scan issues) and Jinja template updates across variants.

Reading public agentic deltas

Public agentic deltas are often shown for 31B and E4B (BFCL, TB2, Tau2 Retail/Airline/Telecom), including figures such as roughly +10.1% on Tau2 Telecom for 31B.

Those tables do not prove identical gains on every size. Prefer your own success and completion rates on production tool schemas.

---

Minimal post-update smoke test

1. Start with one tiny tool (for example fixed JSON or time).
2. Confirm a structured function call is produced.
3. Return the tool result in the tool role.
4. Confirm the final answer is not truncated on a longer generation.

---

Sources

• THE DECODER
• explainx
• AlphaSignal

For separating instructions, guards, and verification in agent loops, see safe harness design for agent loops.

---

How to re-pull weights—and when you must

Because there is no version bump, cached weights do not refresh themselves.

Self-hosted and local runtimes should re-pull the relevant google/gemma-4-* repos. The public distribution surface includes google/gemma-4-31B-it and the broader Gemma 4 collection.

Re-pull checklist

• ✓ Confirm the model ID you actually serve (E2B-it / E4B-it / 12B-it / 26B-A4B-it / 31B-it, etc.)
• ✓ Re-pull the Hugging Face repo and refresh the local cache
• ✓ Update chat templates and tokenizer with the same set
• ✓ On Hopper, verify FA4 is enabled (for example vLLM flash_attn v4 path)
• ✓ Run a tool-call smoke test and a longer generation for truncation
• ✓ For vision only, raise max_soft_tokens toward 1120 when OCR resolution needs it (keep 280 if sufficient)

If you only use a managed API, check the provider’s release notes before assuming a local re-pull is required.

---

Sources

• Hugging Face google/gemma-4-31B-it
• TechTimes
• Google AI: Function calling with Gemma 4

---

How to read the agentic benchmark tables

Treat published net gains as directional hints.

Do not copy 31B/E4B Telecom deltas onto E2B or 12B without testing. Even a large Telecom lift will not transfer if your tool schemas, latency budgets, or stop policies differ.

Recommended priority

1. Your task’s tool success and completion rates (regression suite)
2. Truncation rate and retry cost
3. Prefill/TTFT measurements—only on Hopper FA4 paths
4. Public benchmark deltas (context only)

Start with a staging pilot, proceed only after smoke tests and a short regression suite, and keep a roll back path ready—do not deploy from a chart alone.

---

Official function-calling verification steps

After the refresh, validate tool calling with Google’s documented loop in Function calling with Gemma 4 (last updated 2026-06-04 UTC).

Documented four-stage loop

1. Define tools with JSON schema or Python functions (get_json_schema).
2. Model turn via apply_chat_template(..., tools=...) to obtain a structured function call.
3. Developer turn: execute the tool and append the result as a tool-role message.
4. Final response: model produces the user-facing answer.

Documented model IDs include google/gemma-4-E2B-it, E4B-it, 12B-it, 31B-it, and 26B-A4B-it.

Official notes also warn about Automatic vs Manual Schemas for nested structures—plan manual schemas for complex tools.

---

FAQ

Q1. Is this Gemma 4.1?

No. Coverage describes a same-name weight/template refresh, not a new major version label.

---

Q2. Will an RTX 4090 get the 70% speedup?

The headline FA4 speed claim is Hopper-oriented. On Ada, evaluate quality fixes (tool calling, truncation) instead of headline prefill gains.

---

Q3. What should I do first?

1. Confirm model ID
2. Re-pull from Hugging Face
3. Sync templates/tokenizer
4. Tool smoke test
5. If on Hopper, verify FA4

---

Q4. Must vision users raise max_soft_tokens?

No. Keep the default 280 unless higher OCR resolution is required; then expand toward 1120.

---

Q5. Do cloud-API-only users need a local re-pull?

Often no. Confirm the provider’s model revision instead.

---

Author Perspective

Same-name silent refreshes are less about chasing a “70%” headline and more about revision reproducibility and agent regression.

Teams without Hopper should not treat prefill marketing as a KPI. When agents depend on tool calling, template and weight re-pull can matter more than flat speed numbers.

Keep claim-range numbers out of SLOs until your own smoke tests and short regressions pass.

---

Summary and next steps

Three takeaways:

1. Speed is conditional: Hopper + FA4 claim range (~25–70% prefill, up to ~31% TTFT reduction). Do not expect the headline on Ada.
2. Agent quality fixes may be the real win: tool calling, truncation, template reliability.
3. Re-pull is required for cached local weights: no version bump means no automatic upgrade.

Short next-action checklist

• ✓ Confirm whether your GPU is Hopper (if not, drop speed KPIs)
• ✓ Re-pull the HF model you serve
• ✓ Sync templates/tokenizer
• ✓ Smoke-test one tool end-to-end
• ✓ Expand vision soft tokens only when needed

For open-weight Gemma 4 size context and lower-VRAM coding models, see GLM-5.2 and Gemma 4 12B Coder.

For cost and data-boundary tradeoffs when choosing self-host vs external models, see why U.S. developers route toward lower-cost Chinese models.

Related articles:

• GLM-5.2 and Gemma 4 12B Coder: Low-VRAM Open-Source AI Models Rivaling Opus Performance
• Herdr vs tmux for AI coding agents: state sidebar, install, and migration checks
• How to Assess an AI-Native Org: Enabled, Ready, Native Checklist

Related new article:

• GPT-5.6 Prompting: Lean System Prompts, Autonomy Bounds, and PTC Routing – This published update adds current operational context for Gemma 4 Same-Name Update: FA4 Prefill, Tool Fixes, Re-pull Guide.
• Gemini Notebook rename and cloud code execution: Pro rollout checklist – This published update adds current operational context for Gemma 4 Same-Name Update: FA4 Prefill, Tool Fixes, Re-pull Guide.
• LM Studio Bionic Preview: Local Open-Model Agent for Docs and Code – This published update adds current operational context for Gemma 4 Same-Name Update: FA4 Prefill, Tool Fixes, Re-pull Guide.
• Port Frontier Workflows to Opus/Sonnet with Agent Skills – This published update adds current operational context for Gemma 4 Same-Name Update: FA4 Prefill, Tool Fixes, Re-pull Guide.
• Obsidian Visualization Tools: Graph, Canvas, Excalidraw (2026) – This published update adds current operational context for Gemma 4 Same-Name Update: FA4 Prefill, Tool Fixes, Re-pull Guide.
• Kimi K3 Practical Guide: Pricing, Benchmarks, and API Caveats – This published update adds current operational context for Gemma 4 Same-Name Update: FA4 Prefill, Tool Fixes, Re-pull Guide.