Blog

Gemini 4 Neon: architecture and training runs from primary sources

What Google DeepMind actually disclosed about Gemini 4 Neon across its model cards, earnings transcripts, and post-training talks, and what stays closed.

Published Updated Reading ...
Gemini 4 Neon: architecture and training runs from primary sources

Google DeepMind released Gemini 4 Neon on 18 November 2026 under the API model ID gemini-4-neon (often shortened to g4neon). It sits in the middle of the Gemini 4 lineup, between Gemini 4 Hydrogen for fast, low-cost execution and Gemini 4 Carbon for long-horizon reasoning.

Google DeepMind has not published an academic technical report on arXiv for the Gemini 4 family, unlike open-weight releases such as Gemma 4 (arXiv:2607.02770). Primary evidence comes from four places: the official model cards for Gemini 4 Argon, Carbon, and Neon, Google DeepMind launch and safety disclosures, Alphabet’s Q2 2026 earnings transcript, and public technical talks by Gemini post-training leadership.

Every technical statement below is marked Primary-stated (published directly by Google or Google DeepMind) or Not disclosed (absent from primary documentation).


Release facts

ItemValueStatus
Release date18 November 2026Primary-stated
API model IDgemini-4-neonPrimary-stated
Weight accessClosed weights, API and managed surfaces onlyPrimary-stated
Knowledge cutoffMarch 2026 (earlier in some specific domains)Primary-stated
Input price (standard)$3.00 / 1M tokens ($0.15 cached read, 95% discount)Primary-stated
Output price (standard)$15.00 / 1M tokens, thinking tokens includedPrimary-stated
Batch and flex pricing$1.50 / 1M input, $7.50 / 1M outputPrimary-stated
Academic technical reportNone publishedNot disclosed

Model lineage

The Gemini 4 model cards define an explicit inheritance chain rather than an independent pre-training run for each element.

Figure 1. Documented derivation chain across the Gemini 4 model cards. Edge labels use the exact dependency wording published by Google DeepMind.

  1. Gemini 4 Argon (launched 13 October 2026) is the foundational Gemini 4 model. Its model card describes it as an independent frontier model, not a fine-tune or modification of a prior generation.
  2. Gemini 4 Carbon (announced 28 October 2026 for cyber defenders through the Fairwind Program, with general developer and enterprise access on 10 November 2026) is documented as based on Gemini 4 Argon. Its model card points back to Argon for base architecture, data processing, and training infrastructure.
  3. Gemini 4 Neon (released 18 November 2026) is documented as adapted from Gemini 4 Carbon. Its model card points back to Carbon, and transitively to Argon, for core architecture, pre-training data, preprocessing, and hardware. Gemini 4 Hydrogen carries the same “adapted from Gemini 4 Carbon” dependency in its model card.

Google DeepMind does not specify what “adapted from” means at the weight level. Third-party commentary often assumes knowledge distillation or pruning, but primary documentation never confirms whether Neon is distilled, pruned, served with fewer active experts, or produced by a separate post-training run on the same topology.


Architecture

Because Neon inherits Carbon’s and Argon’s base stack, its disclosed architectural specification comes from the Argon and Neon model cards combined.

ComponentPrimary disclosureStatus
Core topologyLarge-scale sparse Mixture-of-Experts (MoE) TransformerPrimary-stated
Routing behaviourDynamic token-level routing activating a subset of parameters per tokenPrimary-stated
MultimodalityNative multimodal input across text, image, audio (speech and general), and videoPrimary-stated
Document inputsPDFs and native Office and OpenDocument formats supported at the API layerPrimary-stated
Context window1,000,000 unified tokens covering input, thinking, and outputPrimary-stated
Output limit64K default synchronous output; up to 1M tokens via background long decodePrimary-stated
Test-time computeFour configurable thinking levels: low, medium, high (default), xhighPrimary-stated
Total and active parametersNone publishedNot disclosed
Expert count and top-k routingNone publishedNot disclosed
Layer depth, hidden size, attention variantNone publishedNot disclosed

The sparse MoE design separates total parameter capacity from per-token compute during serving (see Mixture of Experts for the broader lineage). The four exposed thinking levels place test-time compute under explicit caller control, billing thinking tokens as output tokens (background in Reasoning and Test-Time Compute).


Training stack and pre-training

Primary disclosures on the base training run come from Alphabet’s Q2 2026 earnings call and the Gemini 4 Argon model card.

Figure 2. Public primary-source timeline from pre-training confirmation to the Neon release.

Pre-training scope

On the Alphabet Q2 2026 earnings call (22 July 2026), CEO Sundar Pichai stated that Google had started its “most ambitious pre-training run yet for Gemini 4” and that competing at the next frontier required “much larger base models,” naming coding and agentic coding as explicit focus areas. Two months later, on 24 September 2026, Google DeepMind’s Koray Kavukcuoglu stated on stage that Gemini 4 had moved into post-training and was being tested internally in Antigravity.

Hardware and software

The Argon model card, inherited by Carbon and Neon, specifies the training stack at the platform level:

  • Hardware: Custom Google Tensor Processing Units (TPUs) with High-Bandwidth Memory (HBM), scaled across TPU Pods. Public model cards do not state which specific TPU generation (such as Trillium v6e or Ironwood TPU7x; see our 2026 accelerator comparison) or what chip count ran each stage.
  • Software: JAX for model definition and numerical execution, paired with ML Pathways as the distributed runtime orchestrating multi-host pre-training and large-scale reinforcement learning.

Pre-training corpus and hygiene

The inherited pre-training pipeline covers publicly available web documents, text, code, images, speech and non-speech audio, and video up to a March 2026 knowledge cutoff. Disclosed preprocessing steps include deduplication, robots.txt compliance, general quality filtering, and case-by-case inline safety filtering to remove CSAM, violent content, and explicit material. Dataset token totals and domain mixing ratios are not disclosed.


Post-training and reinforcement learning

The model cards describe a post-training pipeline combining supervised fine-tuning (SFT) and instruction tuning, human-preference alignment, and large-scale reinforcement learning (RL) targeting multi-step reasoning, software engineering, agentic tool execution, and formal problem-solving and theorem-proving.

How Gemini 4 RL runs address saturation and reward hacking

In technical talks on Gemini 4 post-training (including CODE Summit 2026), Melvin Johnson, Vice President of Research at Google DeepMind leading Gemini post-training, described two structural changes driving Gemini 4 RL runs (see Post-Training and Alignment for how this fits the broader RLVR shift):

  • The learnable-prompt bottleneck. As frontier checkpoints improve, prompts the model passes on every rollout yield zero reward variance and zero gradient signal. Post-training teams must continuously source harder unsolved tasks and feed edge-case failures from internal dogfooding across more than 100,000 Google engineers back into the training loop.
  • Rubric-based grading over binary unit tests. Passing a unit test does not verify that a coding agent solved the problem cleanly, and binary pass/fail rewards invite test-suite gaming and result fabrication. Gemini 4 post-training grades full trajectories inside hardened, isolated environments against dense rubrics based on senior-engineer review: checking intermediate reasoning, architectural constraints, scope discipline, and honest execution reporting.

Figure 3. Post-training feedback loop described in Google DeepMind model cards and Gemini post-training leadership talks.

Neon-specific post-training targets and reported metrics

Google DeepMind’s launch disclosures and model card frame Neon’s adaptation around trajectory efficiency, scope discipline, and prompt-injection defence while keeping false refusals low.

Evaluation areaReported resultStatus
Cost efficiency vs CarbonAbout 32% lower cost per task at default settingsPrimary-stated
Under-service reductionReduced by 43% to 56% vs Carbon baselines across effort levelsPrimary-stated
Overreach reductionReduced by 15% to 35% vs Carbon baselines across effort levelsPrimary-stated
Internal agentic safety suite0.09% misaligned action ratePrimary-stated
Agent honesty evaluation99.2% pass ratePrimary-stated
Coding shortcut benchmark2.45% shortcut ratePrimary-stated
False refusals on coding tasks1.98% on Terminal-Bench 4.0; 0% on Terminal-Bench-SciencePrimary-stated
Indirect prompt injection (IPI)Defensive post-training; strong resilience on Gray Swan IPI at highest thinking tierPrimary-stated

Chain-of-thought (CoT) and action-misalignment monitors ran across training and evaluation to catch unauthorised shortcuts or deceptive workarounds during multi-step tool execution.


Frontier safety evaluations

Under Google DeepMind’s Frontier Safety Framework (FSF), Gemini 4 Neon underwent a Material Capability Change (MCC) assessment relative to Gemini 4 Argon, the last family model given a full critical capability assessment.

Risk domainFSF assessment in Neon Model CardDeployment posture
CBRN (biological)Critical Capability Level (CCL) cannot be ruled out (carried forward from Argon)Real-time and foundational mitigations remain deployed
CybersecurityCritical Capability Level (CCL) cannot be ruled out (carried forward from Argon)Mitigations remain deployed; high-end cyber capabilities gated through controlled channels such as the Fairwind Program

The model card notes that Neon showed capability increases on biological reference scenarios and sufficient uplift for high-impact cyberattacks to retain the Argon CCL posture rather than stepping down to a lower risk tier.


What is not disclosed

Primary research leaves the core quantitative internals proprietary. Anyone evaluating Gemini 4 Neon against open-weight models should note that none of the following appear in any Google DeepMind document:

  • Model sizing and routing: Total parameters, active parameters per token, number of MoE experts, shared-expert count, top-k routing formula, layer count, hidden dimension, vocabulary size, and attention layout (such as grouped-query, latent, or local-global window ratios).
  • Training run telemetry: Pre-training token counts, RL step or rollout counts, total FLOP budgets, wall-clock run durations, optimizers, precision formats (BF16, FP8, or FP4), and learning-rate schedules.
  • Exact hardware allocation: Specific TPU chip generation and pod counts assigned to Argon pre-training versus Carbon and Neon post-training.
  • Exact adaptation recipe: Whether “adapted from Gemini 4 Carbon” denotes weight distillation, structural pruning, reduced active capacity, or a separate post-training checkpoint of the same base architecture.

Scope

  • This article covers primary disclosures only: Google DeepMind model cards, Google launch and safety notes, Alphabet’s Q2 2026 earnings transcript, and public remarks by Gemini research leaders, checked on 20 November 2026.
  • All behavioural, honesty, shortcut, and cost-reduction figures above are vendor-stated by Google DeepMind and have not yet been independently reproduced by Veso AI.
  • Third-party leaderboard standings are excluded from this architecture review. For current independent rankings across frontier models, see our AI model matrix.

Primary sources


Veso AI evaluates frontier models on primary evidence and client workloads. Assessing Gemini 4 for your stack? Get in touch.