Google’s internal deployment machine is moving at breakneck speed. Just as developers and tech enthusiasts were getting comfortable analyzing the rollout patterns of previous model revisions, fresh leaks straight from Google’s distribution network reveal that the company’s next generation of artificial intelligence models is already taking shape behind closed doors.
According to latest evidence extracted from official Google Antigravity ACP (Agent Control Plane) Server release archives, Gemini 4 Carbon and Gemini 4 Argon—along with specialized high-performance variants—have officially entered Google’s internal “teamfooding” phase.
For the uninitiated, “dogfooding” (or “teamfooding” in Google’s internal parlance) is the critical testing phase where internal teams actively use pre-release software in their daily workflows before any public preview or API availability. The appearance of these model identifiers signals that Google DeepMind is actively testing these next-gen architectures in real-world scenarios.
Here is a comprehensive look at the leaked identifiers, what the code diffs reveal, and how this fits into Google’s evolving AI landscape.
What the Leaked Antigravity ACP Revision Reveals
The latest discovery stems from an in-place CDN revision drop (v1.4.3 Revision) detected on dl.google.com. The embedded Discovery Doc within agy_acp_server.par grew by 13,725 bytes, expanding the active internal model registry from 459 to 464 active models.
Key internal and teamfood model identifiers surfaced inside cloudcode-pa_prod_google_rest_v1internal.json:
gemini-4-carbon– The core next-generation Gemini 4 Carbon base model.gemini-4-carbon-teamfood– The dedicated internal testing endpoint for Gemini 4 Carbon.carbon-g-medium-tf– A medium-tier internal teamfood build optimized for balanced latency and capacity.gemini-4-argon-teamfood– Internal testing build for the Argon model series within the Gemini 4 family.gemini-4-argon-max-mode-tf– A specialized, high-compute endpoint designed for maximum reasoning capability, deep research, or extended context processing.
Evidence Summary:
- Server Archive: official Google Antigravity ACP Server release archive on dl.google.com
- Internal Path: google3/google/internal/cloud/code/cloudcode-pa_prod_google_rest_v1internal.json
- Revision Delta: embedded Discovery Doc grew from 1,482,296 bytes to 1,496,021 bytes (+13,725 bytes)
Both Carbon and Argon Max Mode are now actively live internally, giving Google engineers early access to test code generation, multimodal reasoning, and system integrations.
Decoding Google’s Naming Strategy: Carbon and Argon
Google DeepMind has increasingly used elemental codenames alongside functional descriptors for its internal model iterations. Understanding what these codenames represent gives us valuable insight into how Google structures its model tiers.
| Model Codename | Key Focus & Expected Use Case | Operational Status |
Gemini 4 Carbon (gemini-4-carbon) | High-efficiency backbone intended for core developer tasks, fast inference, and balanced agentic workflows. | Active Internal Teamfooding |
Carbon G Medium TF (carbon-g-medium-tf) | Scaled-down intermediate build for daily developer tool integration (e.g., Cloud Code, IDE extensions). | Active Internal Teamfooding |
Gemini 4 Argon (gemini-4-argon-teamfood) | High-intelligence reasoning framework, likely focusing on complex math, logic, and multimodal processing. | Active Internal Teamfooding |
Argon Max Mode (gemini-4-argon-max-mode-tf) | Maximum compute performance endpoint, targeted at deep thinking, long-context analysis, and heavy multi-step task execution. | Active Internal Teamfooding |
The introduction of an explicit “Max Mode” variant (gemini-4-argon-max-mode-tf) is particularly notable. It suggests Google is formalizing a tier specifically tuned for maximum parameters or unconstrained reasoning cycles—similar to how advanced reasoning models leverage extended test-time compute to solve complex technical problems.
If you have been keeping up with our ongoing coverage on the platform, check out our previous breaking report on Google’s Opus Killer: Gemini 4 Carbon Early Access Sparks AI Race, as well as our deep-dive analysis on Gemini 4 Argon: Google’s New Frontier AI Model Explained to see how these architectures developed prior to this internal teamfooding drop.
Why Internal “Teamfooding” Matters for the Public Timeline
When model identifiers hit the internal google3 codebase and cloud production JSON files, it usually indicates that model architecture design and initial pre-training are complete.
System Integration Testing: During teamfooding, Google employees across Cloud Code, Workspace, and internal dev environments test the model for edge cases, hallucination rates, and coding efficiency.
Infrastructure Stress Testing: Serving models at scale requires robust infrastructure. Internal deployments allow Google’s Cloud teams to optimize load balancing and latency.
Refining Safety & Alignment: Fine-tuning and post-training feedback loops are actively collected from real user queries within the company.
Historically, once a model series enters teamfooding across internal CDN endpoints, public experimental releases or developer previews follow within a matter of weeks to months, depending on stability and evaluation benchmarks.
Looking Ahead
Google’s rapid iteration cycle shows no signs of slowing down. With Gemini 4 Carbon and Gemini 4 Argon Max Mode now officially live in internal testing environments, developers can expect significant upgrades in speed, context window handling, and reasoning depth in upcoming public releases.
Stay tuned as we continue to track CDN drops, discovery doc diffs, and official announcements across the AI ecosystem!
