Change
e9ee819
e9ee81982427098ecc9aa4f76bc88e3b5c42da6b · commit on GitHub
gcloud-blog-feed: changed (461586 bytes, HTTP 200)
raw/gcloud-blog-feed/response.xml modified
- Source
- gcloud-blog-feed
- Lines added
- +104
- Lines removed
- -335
- Stored bytes at this commit
- 461,586
- Timestamp
- observed
- Raw artifact at this commit
- raw/gcloud-blog-feed/response.xml
Recorded headers
| observed_at | 2026-10-28T05:42:13.646Z |
|---|---|
| origin_date | null |
| status | 200 |
| final URL | https://cloudblog.withgoogle.com/products/ai-machine-learning/rss/ |
| etag | null |
| last-modified | null |
| date | Wed, 28 Oct 2026 05:42:13 GMT |
| age | null |
| cache-control | public, max-age=600, s-maxage=600 |
| cf-cache-status | null |
| content-encoding | gzip |
| content-length | 126609 |
@
@@ -1,5 +1,102 @@<?xml version="1.0" encoding="utf-8"?>-
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:media="http://search.yahoo.com/mrss/"><channel><title>AI & Machine Learning</title><link>https://cloud.google.com/blog/products/a…+
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:media="http://search.yahoo.com/mrss/"><channel><title>AI & Machine Learning</title><link>https://cloud.google.com/blog/products/a…+
<p><span style="vertical-align: baseline;">To help you get ahead of agentic-era Identity and Access Management, today we’re introducing IAM Helper. This suite of intelligent, agent-facing building blocks can automate complex, end-to-end IAM workflows. </span></p>+
<p><span style="vertical-align: baseline;">As an agent-native orchestration and interface layer for Google Cloud IAM, it augments Google Cloud’s core IAM engines — including Policy Intelligence and Privileged Access Manager (PAM) — making IAM skills callable directly by autonomous code a…+
<h3><strong style="vertical-align: baseline;">IAM Helper Agent Skills: Fast, secure deployment with automated governance</strong></h3>+
<p><span style="vertical-align: baseline;">Manual troubleshooting of permission blockers creates friction for engineering and administrators, while shortcuts can lead to broad agent over-permissioning that puts least privilege at risk. </span></p>+
<p><span style="vertical-align: baseline;">IAM Helper empowers agents to handle access management dynamically. It shifts the burden of policy evaluation away from human administrators and LLMs, offloading complex orchestration directly to the Google Cloud backend, to maximize token effic…+
<p><span style="vertical-align: baseline;">Rather than operating as isolated utilities, IAM Helper Agent Skills guide agents through the cloud access lifecycle. These end-to-end workflows, helper scripts, and guardrails packaged using the open </span><a href="https://agentskills…+
<h3><strong style="vertical-align: baseline;">Built-in guardrails and real-world impact</strong></h3>+
<p><span style="vertical-align: baseline;">Every IAM Helper agent skill separates read-only discovery from policy changes: Diagnostics and simulations run automatically, while any change to access policies pauses for human-in-the-loop (HITL) confirmation based on customizable risk tiers.…+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Scaled user-assisted remediation</strong><span style="vertical-align: baseline;">: Suppose a developer's coding agent hits a 403 error while deploying a new application. Rather than filing a support ticket and i…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Autonomous pipeline self-healing</strong><span style="vertical-align: baseline;">: An infrastructure agent encounters a permission block during a routine deployment pipeline. Rather than failing the build, the p…+
</li>+
</ul>+
<p><span style="vertical-align: baseline;">IAM Helper Agent Skills are designed for maximum token and cost efficiency, allowing agents to execute multi-step administrative operations autonomously.</span></p>+
<ol>+
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Authoring compliant, least-privilege policies</strong><span style="vertical-align: baseline;">: As DevOps engineers translate access requirements into JSON, YAML, or Terraform HCL, iam-helper-for-policy-manageme…+
</li>+
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Unblocking 403 errors in seconds</strong><span style="vertical-align: baseline;">: When an on-call engineer or deployment agent hits a cryptic 403 PERMISSION_DENIED error, iam-helper-for-troubleshooting identifi…+
</li>+
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Simulating blast radius before rolling out changes</strong><span style="vertical-align: baseline;">: Before a platform engineer updates a policy or applies IAM Recommender role reductions, they need confidence t…+
</li>+
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Elevating access just-in-time instead of granting standing privileges</strong><span style="vertical-align: baseline;">: When responders or maintenance agents need temporary production access, iam-helper-for-priv…+
</li>+
</ol>+
<p><span style="vertical-align: baseline;">While IAM Helper Agent Skills supply the high-level workflow logic, they are all based on a common set of foundational MCP Tools. </span></p>+
<h3><strong style="vertical-align: baseline;">Powered by Google Cloud IAM MCP tools</strong></h3>+
<p><span style="vertical-align: baseline;">Google Cloud IAM Helper MCP tools provide the backend connection through atomic, read-only functions exposed via Google-managed MCP servers across core IAM endpoints (Policy Troubleshooter, Policy Analyzer, Policy Simulator, and IAM v1/v2 APIs).…+
<p><span style="vertical-align: baseline;">Rather than requiring teams to write custom API wrappers, MCP tools give agents a standard way to query the cloud backend and get exact answers without parsing verbose policy documents in the prompt, maximizing token efficiency. Agents can dynam…+
<p><span style="vertical-align: baseline;">Because operations like hierarchy traversal, group expansion, and policy evaluation happen on dedicated IAM endpoints (spanning IAM v1 allow roles, IAM v2 deny, Policy Troubleshooter, and Policy Analyzer), skills run quickly and accurately with …+
<h3><strong style="vertical-align: baseline;">Getting started</strong></h3>+
<p><span style="vertical-align: baseline;">Accelerate your agentic workflows, reduce operational overhead, and bring intelligent, scalable security to your deployment pipelines. Start building with IAM Helper Agent skills today by exploring our </span><a href="https://github.com…+
<p><span style="vertical-align: baseline;">Today, we are bringing this same frontier embodied reasoning model to enterprise production environments: </span><strong style="vertical-align: baseline;">Gemini Robotics ER 2 is now available in public preview on the Gemini platform…+
<h2><strong style="vertical-align: baseline;">Powering physical agents for enterprise robotics</strong></h2>+
<p><span style="vertical-align: baseline;">Gemini Robotics ER 2 transforms a robot into a physical agent. The model combines deep spatio-temporal reasoning and long-horizon planning, enabling robots to orchestrate steps to complete a task, self-correct, and generalize to novel situations…+
<p><span style="vertical-align: baseline;">Here is what the model delivers for businesses building and operating robotics:</span></p>+
<h3><strong style="vertical-align: baseline;">1. Physical task verification, understanding, and self-correction</strong></h3>+
<p><span style="vertical-align: baseline;">In physical operations, a robot needs to know whether a step actually succeeded before moving on to the next one. Gemini Robotics ER 2 adapts to new and unfamiliar situations on the fly by reasoning over continuous video feeds. The model:</sp…+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><span style="vertical-align: baseline;">Classifies task progress into binary status, discrete categories, or continuous scores to improve real-time situational awareness and enable real-time adjustment or iteration on failed steps, without restarting an entire work…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><span style="vertical-align: baseline;">Pinpoints the exact video frame where a critical event occurs (such as when to stop pouring a liquid) with 91.3% accuracy and a 0.96-second mean absolute frame distance, running at 4x the execution speed and a fraction of the…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><span style="vertical-align: baseline;">It advances our core spatial reasoning capability, including visual question answering and instrument reading. For example, it can read and interpret values across 10+ instrument categories including circular dials, sight gla…+
</li>+
</ul>+
<h3><strong style="vertical-align: baseline;">2. One intelligence layer across any embodiments</strong></h3>+
<p><span style="vertical-align: baseline;">Gemini Robotics ER 2 orchestrates robots of many shapes and sizes, scaling its intelligence from stationary arms and wheeled rovers to quadrupeds and complex forms. It also unlocks multi-robot collaboration, allowing different types of robots to…+
<h3><strong style="vertical-align: baseline;">3. Agentic tool orchestration and natural language interaction</strong></h3>+
<p><span style="vertical-align: baseline;">Gemini Robotics ER 2 thinks about what comes next while simultaneously orchestrating actions. Developers can declare low-level control interfaces — such as VLA models, navigation APIs, or custom functions — as tools across simulation, real-worl…+
<h3><span style="vertical-align: baseline;">4. </span><strong style="vertical-align: baseline;">Build safe, adaptive physical agents in Google Cloud</strong></h3>+
<p><span style="vertical-align: baseline;">Developers and enterprise teams use the Gemini platform to simulate, orchestrate, and manage physical agents, which are then deployed to run directly on physical robots at the edge using Google Cloud's scalable infrastructure:</span></p…+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Simulate fast: </strong><span style="vertical-align: baseline;"> Test robot behaviors before touching hardware by running simulations in familiar environments like MuJoCo-Warp and NVIDIA Isaac Sim. Then generate…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Train smart</strong><span style="vertical-align: baseline;">: Use Gemini Robotics ER 2's spatial pointing, moment finding, and progress classification to curate and label demonstration videos in Cloud Storage an…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Operate easily: </strong><span style="vertical-align: baseline;">Deploy Gemini Robotics ER 2 workflows on the Gemini platform with enterprise IAM, security controls, and built-in function calling. Offload latenc…+
</li>+
</ul>+
<h2><strong style="vertical-align: baseline;">How customers are building with Gemini Robotics ER 2 in Gemini platform</strong></h2>+
<p><em><span style="vertical-align: baseline;">“We’ve seen exponential growth in customers using AIVI-Learning every month, and Gemini Robotics ER 2 brings them even more value. Video changes what an inspection can be. Spot isn’t just checking whether something looks right in a mom…+
<p><span style="vertical-align: baseline;">Marco da Silva, VP and General Manager of Spot at </span><strong style="vertical-align: baseline;">Boston Dynamics</strong></p>+
<p><span style="vertical-align: baseline;"><span style="font-style: italic; vertical-align: baseline;">“DeepReach AI is the real-world skills engine for physical AI. Our hardware and AI platform capture how skilled people work all over the world, and turn it into data flywheel for …+
<p><span style="vertical-align: baseline;"><span style="vertical-align: baseline;">Chris Liu, Co-founder and CTO at <strong>DeepReach</strong></span></span></p>+
<p><span style="font-style: italic; vertical-align: baseline;">“Advanced reasoning models have the potential to help robots interpret complex environments, plan actions and adapt to changing conditions, opening new possibilities for physical AI. Google’s Gemini Robotics ER 2 model repres…+
<p><span style="vertical-align: baseline;"><span style="vertical-align: baseline;">Barrie Mullins, Senior Director of Arm Total Design for Physical AI, at <strong>Arm</strong></span></span></p>+
<p><span style="font-style: italic; vertical-align: baseline;">“We aim to accelerate materials R&amp;D workflows so our scientists can test more ideas, faster. Most lab automation today relies on rigid scripts for single, isolated tasks. By using Gemini Robotics ER 2, we are giving o…+
<p><span style="vertical-align: baseline;">Prof. Loh Xian Jun, Executive Director at <strong>ASTAR Institute of Materials Research and Engineering (ASTAR IMRE)</strong> and Executive Director of MARVEL-AI”</span></p>+
<h2><strong style="vertical-align: baseline;">Get started with Gemini Robotics ER 2 in Gemini platform</strong></h2>+
<p><span style="vertical-align: baseline;">Developers and enterprise teams can start building with Gemini Robotics ER 2 today in preview via Model Garden in Gemini platform.</span></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Build in Gemini platform:</strong><span style="vertical-align: baseline;"> Visit the</span><a href="https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/gemini-robotics-er">&l…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Explore the Developer Guides:</strong><span style="vertical-align: baseline;"> Check out sample code and bidirectional streaming guides in the</span><a href="https://ai.google.dev/gemini-api/docs/roboti…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Read the Research:</strong><span style="vertical-align: baseline;"> Dive into the benchmarks and safety technical report in the</span><a href="https://blog.google/innovation-and-ai/models-and-research/g…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Explore Physical AI on Google Cloud:</strong><span style="vertical-align: baseline;"> Visit</span><a href="https://cloud.google.com/solutions/physical-ai"><span style="vertical-align: baseline;"&g…+
</li>+
</ul></div></description><pubDate>Tue, 27 Oct 2026 13:00:00 +0000</pubDate><guid>https://cloud.google.com/blog/products/ai-machine-learning/gemini-robotics-embodied-reasoning-2-is-now-available-on-google-cloud/</guid><category>AI & Machine Learning</category><media:content height="54…<div class="block-image_full_width">@
@@ -211,12 +308,12 @@<p><span style="vertical-align: baseline;">No matter how you choose to deploy, you can access Anthropic's latest frontier models on Google Cloud, including </span><a href="https://console.cloud.google.com/agent-platform/publishers/anthropic/model-garden/claude-fable-5-1"><…<div class="block-code"><dl> <dt>code_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;---\r\nname: test-triager\r\ndescription: Runs unit tests, classifies failures, and returns a concise summary of failing assertions.\r\nmodel: flash\r\ntools:\r\n - read\r\n - edit\r\n---\r\n\r\nRun the requested t…+
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;---\r\nname: test-triager\r\ndescription: Runs unit tests, classifies failures, and returns a concise summary of failing assertions.\r\nmodel: flash\r\ntools:\r\n - read\r\n - edit\r\n---\r\n\r\nRun the requested t…</dl></div><div class="block-paragraph"><p data-block-key="5p921">And a cross-model code reviewer subagent in <code>.agents/agents/gemini-judge.md</code>:</p></div><div class="block-code"><dl> <dt>code_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &quot;---\r\nname:cross-model-judge\r\ndescription: Independent LLM-as-a-judge verifier that audits diffs and edge cases using Gemini Pro.\r\nmodel: pro\r\ntools:\r\n - read\r\n---\r\n\r\nReview the current workspace diff…+
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &quot;---\r\nname:cross-model-judge\r\ndescription: Independent LLM-as-a-judge verifier that audits diffs and edge cases using Gemini Pro.\r\nmodel: pro\r\ntools:\r\n - read\r\n---\r\n\r\nReview the current workspace diff…</dl></div><div class="block-paragraph"><p data-block-key="5p921">With this setup, you can delegate rapid test-and-classify loops to <code>test-triager</code> and trigger <code>cross-model-judge</code> to validate the final diff, combining the strengths of each model in a si…<div class="block-video">@
@@ -1405,7 +1502,7 @@<p><code style="vertical-align: baseline;">AI.COUNT_TOKENS</code><span style="vertical-align: baseline;"> returns the number of tokens a model's tokenizer produces for a text string, without making a billable generation call. Run it over a column to measure your input data:&l…<div class="block-code"><dl> <dt>code_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &quot;-- Count the input text tokens for each product review\r\nSELECT\r\n SUM(\r\n AI.COUNT_TOKENS(\r\n CONCAT(&#x27;Classify the sentiment of this review as positive, negative, or neutral: &#x27;, review…+
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &quot;-- Count the input text tokens for each product review\r\nSELECT\r\n SUM(\r\n AI.COUNT_TOKENS(\r\n CONCAT(&#x27;Classify the sentiment of this review as positive, negative, or neutral: &#x27;, review…</dl></div><div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">The count is a lower bound on input tokens. It doesn't include output or thinking tokens, the instructions each AI function adds around your input, or non-text inputs such as images. Use it to:</spa…<ul>@
@@ -1432,7 +1529,7 @@<p><span style="vertical-align: baseline;">If a small sample run shows an average </span><strong style="vertical-align: baseline;">100:1 input-to-output token ratio</strong><span style="vertical-align: baseline;">, you can solve for a daily budget:</span><…<div class="block-code"><dl> <dt>code_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;Cost: 0.125 × I + 0.75 × O = 100 (I, O in millions of tokens)\r\nRatio: I = 100 × O&#x27;), (&#x27;language&#x27;, &#x27;&#x27;), (&#x27;caption&#x27;, &lt;wagtail.rich_text.Ric…+
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;Cost: 0.125 × I + 0.75 × O = 100 (I, O in millions of tokens)\r\nRatio: I = 100 × O&#x27;), (&#x27;language&#x27;, &#x27;&#x27;), (&#x27;caption&#x27;, &lt;wagtail.rich_text.Ric…</dl></div><div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">Solving gives about </span><strong style="vertical-align: baseline;">754 million input tokens</strong><span style="vertical-align: baseline;"> and </span><strong style=…<h3><span style="vertical-align: baseline;">Best practices for managing AI cost</span></h3>@
@@ -2442,7 +2539,7 @@<p><span style="vertical-align: baseline;">Integrating cloud management into your agents no longer requires packaging Google Cloud CLI binaries, managing local execution runtimes, or maintaining dependencies inside agent container images. Because the Google Cloud CLI remote MCP server im…<div class="block-code"><dl> <dt>code_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;{\r\n &quot;mcpServers&quot;: {\r\n &quot;google-cloud-cli&quot;: {\r\n &quot;uri&quot;: &quot;https://cloudcli.googleapis.com/mcp&quot;,\r\n ...\r\n }\r\n }\r\n}&…+
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;{\r\n &quot;mcpServers&quot;: {\r\n &quot;google-cloud-cli&quot;: {\r\n &quot;uri&quot;: &quot;https://cloudcli.googleapis.com/mcp&quot;,\r\n ...\r\n }\r\n }\r\n}&…</dl></div><div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">Your agent immediately gains access to execute gcloud and bq commands in a secure, network-isolated cloud sandbox.</span></p><p><span style="vertical-align: baseline;">Authentication is handled via keyless Agent Identity for hosted Google Cloud platforms, or standard OAuth 2.0 for external runtimes. For authentication options, see the </span><a href="https://docs.cloud.google.com/mcp/set-up-authentica…@
@@ -2579,332 +2676,4 @@<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;"><p role="presentation"><a href="https://www.youtube.com/watch?v=-fb0ycu4kiU" rel="noopener" target="_blank"><span data-rich-links='{"fple-t":"Automate Google Cloud with Cloud CLI Remote MCP Server","fple-u":"https://www.youtube.com/watch?v=-fb0ycu4kiU","fple-mt":null,"type":"first-par…</li>-
</ul></div></description><pubDate>Wed, 30 Sep 2026 16:00:00 +0000</pubDate><guid>https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-cli-remote-mcp-server-in-preview/</guid><category>Data Analytics</category><category>Developers & Practitioners</category><category…-
<p><span style="vertical-align: baseline;">Yet as startup architectures mature, a clear divide has emerged between teams struggling with margins and those scaling sustainably. The most effective engineering teams have abandoned the one-size-fits-all model strategy.</span></p>-
<p><span style="vertical-align: baseline;">In the early days of LLMs the default architecture was simple: send every interaction to the largest model available. But as applications move into production, serving millions of people and running autonomous multi-agent workflows, relying on a…-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Latency penalties: </strong><span style="vertical-align: baseline;">Relying entirely on cloud round trips makes it difficult to deliver the sub-second responsiveness that interactive mobile and desktop apps requ…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Infrastructure overhead: </strong><span style="vertical-align: baseline;">Self-hosting large open models with more than 70 billion parameters forces early-stage teams to act like infrastructure providers, pullin…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Margin erosion:</strong><span style="vertical-align: baseline;"> Sending high-frequency, structured tasks (like intent routing, JSON extraction, or status validation) to general-purpose frontier endpoints spends…-
</li>-
</ul>-
<p><span style="vertical-align: baseline;">Great engineering teams pick the right tool for each job. Most production requests don’t require a frontier generalist, and routing every call to one can actually slow your product down. Instead, the winning pattern is a compound AI stack: pairi…-
<p><span style="vertical-align: baseline;">It’s for these reasons that an open model like Gemma belongs in your model lineup. With more than one billion downloads across the </span><a href="https://deepmind.google/models/gemma/gemmaverse/" rel="noopener" target="_blank"><s…-
<h3><span style="vertical-align: baseline;">Built under one roof</span></h3>-
<p><span style="vertical-align: baseline;">Gemma is built by Google DeepMind using the same foundational research and architecture advances behind the Gemini models. Because they share common DNA and developer tooling, your team can prototype in Google AI Studio and design hybrid archite…-
<p><span style="vertical-align: baseline;">Released under a commercially permissive </span><strong style="vertical-align: baseline;">Apache 2.0 license</strong><span style="vertical-align: baseline;">, </span><a href="https://ai.google.dev/gemma/docs/core…-
<h3><span style="vertical-align: baseline;">Real proof: How startups are winning with Gemma</span></h3>-
<p><span style="vertical-align: baseline;">Founders are using Gemma to solve urgent problems around unit economics, output accuracy, and responsiveness.</span></p>-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Flipping the architecture:</strong><span style="vertical-align: baseline;"> </span><a href="https://heycue.io/" rel="noopener" target="_blank"><span style="text-decoration: underline; vertical-ali…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">True edge independence: </strong><span style="vertical-align: baseline;">Mobile development studio </span><a href="https://hubx.co/" rel="noopener" target="_blank"><span style="text-decoration: un…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Scientific discovery and air-gapped security</strong><span style="vertical-align: baseline;">: K-Dense has built </span><a href="https://www.k-dense.ai/products/faraday" rel="noopener" target="_blank"&g…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Unlocking infinite gameplay and retention:</strong><span style="vertical-align: baseline;"> Gaming company</span><a href="https://aidungeon.com/" rel="noopener" target="_blank"><span style="vertic…-
</li>-
</ul>-
<h3><span style="vertical-align: baseline;">Four workloads where Gemma wins for startups</span></h3>-
<p><span style="vertical-align: baseline;">If you’re evaluating where Gemma fits into your stack today, start with these four jobs:</span></p>-
<h4><span style="vertical-align: baseline;">1. Edge and local execution (low latency, true privacy)</span></h4>-
<p><span style="vertical-align: baseline;">If you’re building mobile apps, developer desktop tools, robotics, or offline-first experiences, every cloud round-trip adds latency that people can feel. Gemma can run directly on laptops (including Apple silicon), smartphones, and local applia…-
<p><span style="vertical-align: baseline;">You can handle many local interactions on-device for zero incremental cost, and keep a bridge to frontier models in the cloud for the requests that need it. When a local workflow calls for large-scale multimodal reasoning, long-context data synt…-
<h4><span style="vertical-align: baseline;">2. High-throughput triage and agent routing</span></h4>-
<p><span style="vertical-align: baseline;">In multi-agent architectures, agents spend a surprising amount of tokens on simple tasks like checking statuses, classifying intent, and routing tickets. With Gemma as your front-line gatekeeper, those high-volume background tasks run on a compa…-
<h4><span style="vertical-align: baseline;">3. Task-specific fine-tuning for real moats</span></h4>-
<p><span style="vertical-align: baseline;">Adapting a model to your proprietary data is one way to build a competitive moat. Because Gemma gives you full access to model weights and has a compact memory footprint, your team can run parameter-efficient fine-tuning (LoRA or QLoRA) on a sin…-
<h4><span style="vertical-align: baseline;">4. Turnkey vertical starting lines</span></h4>-
<p><span style="vertical-align: baseline;">DeepMind releases domain-specific variants of Gemma, so you don’t have to start from scratch. One example is </span><a href="https://research.google/blog/medgemma-our-most-capable-open-models-for-health-ai-development/" rel="noopener" t…-
<p><span style="vertical-align: baseline;">Beyond healthcare, biotech startups use </span><a href="https://research.google/blog/teaching-machines-the-language-of-biology-scaling-large-language-models-for-next-generation-single-cell-analysis/" rel="noopener" target="_blank">&l…-
<h3><span style="vertical-align: baseline;">Deploy wherever your business lives</span></h3>-
<p><span style="vertical-align: baseline;">Gemma is designed to fit into your existing engineering stack without lock-in:</span></p>-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Apache 2.0 licensing</strong><span style="vertical-align: baseline;">: Gemma 4 ships under the Apache 2.0 license, giving startups the freedom to fine-tune, quantize, redistribute, and deploy commercial products…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Day-zero open tooling</strong><span style="vertical-align: baseline;">: Run and fine-tune Gemma with the tools your engineers already use, including vLLM, Ollama, llama.cpp, LM Studio, MLX, Unsloth, Hugging Face…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Serverless and managed cloud deployment</strong><span style="vertical-align: baseline;">: Prototype immediately in </span><a href="https://aistudio.google.com/" rel="noopener" target="_blank"><spa…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Enterprise-ready safety</strong><span style="vertical-align: baseline;">: Gemma undergoes rigorous pre-release safety evaluations, data filtering, and red-teaming, and pairs with ShieldGemma 2 to help you meet e…-
</li>-
</ul>-
<h3><span style="vertical-align: baseline;">Build with Gemma: What to do this week</span></h3>-
<p><span style="vertical-align: baseline;">Great technical architecture isn't about finding one model to do everything. It’s about assembling the right tool for each job so you can move faster, protect your runway, and ship a superior product.</span></p>-
<p><span style="vertical-align: baseline;">Here’s my challenge to your engineering team this week:</span></p>-
<ol>-
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Audit your model calls:</strong><span style="vertical-align: baseline;"> Look at your logging dashboard and identify three high-volume, deterministic tasks (such as intent classification, JSON validation, or sum…-
</li>-
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Benchmark Gemma:</strong><span style="vertical-align: baseline;"> Run a quick test with a compact Gemma model locally or on a single endpoint. Measure the latency and calculate what happens to your gross margins…-
</li>-
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Redirect your runway:</strong><span style="vertical-align: baseline;"> Take the capital and engineering hours you save on compute and invest them back into your core differentiators.</span></p>-
</li>-
</ol>-
<p><span style="vertical-align: baseline;">You can download the Gemma weights directly or deploy them through Model Garden. If you need compute credits and technical architecture reviews to get up and running, the Google for Startups team is ready to help you build - </span><a h…-
<p><span style="vertical-align: baseline;">Now, you can adapt Gemini with the service — teaching the model from a reward signal you define, rather than from a fixed set of labeled answers. This unlocks a class of problems that supervised fine-tuning (SFT) struggles with: tasks that are h…-
<p><span style="vertical-align: baseline;">In this guide, we will walk through practical best practices for using RL fine-tuning service. We'll start with a short tour of the RL training loop, how to decide if and when to use RL, and introduce how to get the most value from this approach…-
<h3><span style="vertical-align: baseline;">What is RLFT? </span></h3>-
<p><span style="vertical-align: baseline;">RLFT adapts Gemini from a </span><strong style="vertical-align: baseline;">reward signal you define</strong><span style="vertical-align: baseline;"> rather than labeled answers. Instead of authoring a large set of gold ex…-
<div class="block-image_full_width">-
-
-
-
-
-
-
-
<div class="article-module h-c-page">-
<div class="h-c-grid">-
-
-
<figure class="article-image--large-
-
-
h-c-grid__col-
h-c-grid__col--6 h-c-grid__col--offset-3-
-
-
"-
>-
-
-
-
-
<img-
src="https://storage.googleapis.com/gweb-cloudblog-publish/images/1_-_Single-Step_RL_Training_Loop.max-1000x1000.png"-
-
alt="1 - Single-Step RL Training Loop">-
-
</a>-
-
</figure>-
-
-
</div>-
</div>-
-
-
-
-
-
</div>-
<div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">At each training step the service generates multiple candidate responses to your prompts, scores them with your reward, and improves the model so that higher-scoring responses become more likely while …-
<p><span style="vertical-align: baseline;">Three properties define what RLFT can and can't do:</span></p>-
<ol>-
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">-
<p role="presentation"><span style="vertical-align: baseline;">It </span><strong style="vertical-align: baseline;">learns from the model's own outputs:</strong><span style="vertical-align: baseline;"> It refines what the model already produces rather than copying …-
</li>-
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">-
<p role="presentation"><span style="vertical-align: baseline;">It </span><strong style="vertical-align: baseline;">rewards outcomes, not paths:</strong><span style="vertical-align: baseline;"> Any response that reaches a good result earns reward, which fits open…-
</li>-
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">-
<p role="presentation"><span style="vertical-align: baseline;">It </span><strong style="vertical-align: baseline;">amplifies existing competence: </strong><span style="vertical-align: baseline;"> It makes </span><span style="font-style: italic; vertical-…-
</li>-
</ol>-
<h3><span style="vertical-align: baseline;">When to use RLFT</span></h3></div>-
<div class="block-image_full_width">-
-
-
-
-
-
-
-
<div class="article-module h-c-page">-
<div class="h-c-grid">-
-
-
<figure class="article-image--large-
-
-
h-c-grid__col-
h-c-grid__col--6 h-c-grid__col--offset-3-
-
-
"-
>-
-
-
-
-
<img-
src="https://storage.googleapis.com/gweb-cloudblog-publish/images/2_-_RLFT_Approaches_-_Direct_RL_or_SFT_War.max-1000x1000.png"-
-
alt="2 - RLFT Approaches - Direct RL or SFT Warmup to Continuous RLFT">-
-
</a>-
-
</figure>-
-
-
</div>-
</div>-
-
-
-
-
-
</div>-
<div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">Prompting and SFT handle most adaptation; exhaust them first. RLFT earns its keep when you can </span><strong style="vertical-align: baseline;">grade a response but can't cheaply author it&…-
<p><span style="vertical-align: baseline;">SFT and RLFT are complementary, not competing:</span></p>-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Direct RLFT</strong><span style="vertical-align: baseline;"> when the base model already succeeds part of the time — enough for the reward to tell better answers from worse ones.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Two-stage SFT → RLFT</strong><span style="vertical-align: baseline;"> when you have SFT data or the base success rate is too low for RL to gain traction. Use SFT as a short, cheap warm start — kept light, since …-
</li>-
</ul>-
<p><span style="vertical-align: baseline;">Across early adopters, these patterns show where RLFT delivers the most value — each scoring an outcome the business cares about but could never cheaply demonstrate.</span></p>-
<h3><span style="vertical-align: baseline;">Use cases for RLFT</span></h3>-
<p><strong style="vertical-align: baseline;">AI-powered NPCs in games</strong></p>-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> In-character, on-brand dialogue held across long, multilingual, multi-turn conversations.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> Off-the-shelf models break immersion — wrong language, hallucinated items, ignored players, repetitive loops.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Objective and reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">Gemini autorater (LLM-as-a-judge)</strong><span style="vertical-alig…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> Loops and language drift disappeared and state syntax held, making shippable in-game characters viable at scale.</span></p>-
</li>-
</ul>-
<p><strong style="vertical-align: baseline;">Structured entity extraction</strong></p>-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> Pulling a set of items from unstructured documents, such as supplier invoices and shipping manifests, into structured records automatically</span><…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> The long tail where SFT plateaus — missing required fields (recall) or inventing ones that aren't there (precision).</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Objective and reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">rule-based precision/recall reward</strong><span style="vertical-ali…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> Field-level accuracy rose on noisy real-world documents where tuning had stalled, turning a manual review step into an automated one.</span></p&…-
</li>-
</ul>-
<p><strong style="vertical-align: baseline;">Content moderation</strong></p>-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> Applying intricate policies and decision trees at scale.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> Models hallucinate false positives or reward-hack with invalid formats to dodge evaluation.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Objective and reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">Cloud Run reward</strong><span style="vertical-align: baseline;">…-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> The model handled complex exemption carve-outs, sharply cut false positives, and stopped reward hacking — reducing the human-escalation volume that makes…-
</li>-
</ul>-
<p><strong style="vertical-align: baseline;">Code measured by execution</strong></p>-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> SQL or API calls graded on whether they actually run against customer data.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> SFT mimics one reference query and breaks on unseen proprietary schemas.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Objective &amp; Reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">code-execution reward</strong><span style="vertical-align: bas…-
</li>Diff display stops at 400 lines. The line counts above are from the whole diff. 98 lines shown here cut at 300 characters. The raw artifact at this commit is linked above.