AI Architecture Shifts: TabFM, Inkling, and Vera Rubin July

Three distinct architecture advances landed in July 2026 that collectively reframe how AI models are built, served, and refined. Google Research published TabFM, a zero-shot foundation model for tabular data. Thinking Machines released Inkling, a 975-billion-parameter open-weights multimodal mixture-of-experts model. And NVIDIA detailed why its Vera Rubin platform targets intelligence per dollar as the defining metric for agentic post-training workloads.

Google’s TabFM Eliminates Per-Dataset Training

Google Research’s TabFM generates predictions for a new, unseen table in a single forward pass — no retraining, no hyperparameter search, no feature engineering pipeline. The model treats tabular prediction as an in-context learning problem, the same approach that made large language models effective at zero-shot text tasks. For enterprise teams, that compresses time-to-production from weeks of pipeline work to a single API call.

Traditional gradient-boosted tree workflows require data scientists to clean inputs, impute missing values, encode categorical variables, and run repetitive hyperparameter optimization loops across learning rates, tree depths, and regularization grids. After deployment, the operational burden continues. According to VentureBeat’s coverage, Weihao Kong, Research Scientist at Google Research, told VentureBeat that traditional models “incur ongoing operational debt through data drift monitoring and retraining pipelines to stay accurate.”

TabFM sidesteps that entirely. By framing each new table as context rather than a training target, the model generalizes across datasets it has never seen. Google Research’s blog post positions this as bringing the zero-shot inference capability already standard in generative AI to the tabular domain, where the vast majority of business data — warehouses, CRMs, financial ledgers — still lives.

Thinking Machines Open-Sources 975B-Parameter Inkling

Thinking Machines, founded by former OpenAI CTO Mira Murati, released Inkling on July 14, 2026 under an Apache 2.0 license — making it freely usable and modifiable for commercial deployments. At 975 billion total parameters, Inkling is a natively multimodal mixture-of-experts system that reasons across text, images, and audio. Weights are available on Hugging Face and through the company’s own API, Tinker.

On third-party benchmarks, Inkling posts 77.6% on SWE-bench Verified, beating NVIDIA Nemotron 3’s 71.9% on the same software engineering evaluation. Voice understanding scores 91.4% on VoiceBench, compared to 94.4% for Gemini 3.1 Pro on high reasoning effort — competitive but below the current state-of-the-art frontier. According to Thinking Machines’ announcement, the model was designed “to answer directly on topics that may be subject to censorship,” a deliberate design choice targeting enterprises that need factual outputs regardless of topic sensitivity.

A lighter companion model, Inkling-Small at 276 billion parameters, targets workloads where lower inference cost outweighs peak capability. Both models include a “controllable thinking effort” mechanism that lets operators dial compute expenditure per query — a meaningful departure from fixed-cost frontier inference.

The open-source community responded quickly. In a post on X, Thinking Machines co-founder John Schulman shared his perspective on the release, while Horace He, a researcher at Thinking Machines previously from PyTorch, noted on X the engineering complexity involved in building the system.

MoE Architecture Dominates Frontier Serving

Both Inkling and the broader frontier model ecosystem have converged on mixture-of-experts architecture for a concrete reason: it decouples parameter count from per-token compute cost. Only a subset of experts activates per forward pass, which keeps inference costs manageable even as total model size grows into the hundreds of billions of parameters.

According to the NVIDIA AI Blog, virtually every frontier AI model today runs on a MoE architecture. Serving these models efficiently at scale requires GPU domain size — the number of GPUs connected over high-speed scale-up interconnects — to grow accordingly. NVIDIA’s Hopper generation standardized an eight-GPU domain; the company’s Blackwell NVL72 platform scales that to 72 GPUs per domain, which NVIDIA says is necessary to serve large MoE models under production load without throughput bottlenecks.

The core metric NVIDIA emphasizes is performance per watt. Power is a fixed constraint for data center operators: a given facility has a maximum draw, and revenue scales with how many tokens that budget can generate. Performance per watt cannot be inflated by benchmark selection — it reflects real-world throughput under real power limits.

NVIDIA Vera Rubin Targets Post-Training Efficiency

NVIDIA’s Vera Rubin platform reframes the central AI compute workload as post-training rather than initial training or inference alone. The NVIDIA AI Blog’s Vera Rubin post argues that agentic AI has made post-training continuous rather than a one-time finishing step — agents encounter new tools, edge cases, and deployment environments week to week, each requiring model refinement.

The defining metric for this workload is intelligence per dollar: the yield extracted from every forward and backward pass in a continuous learning cycle. Because inference cost is measured in cost per token, every improvement to token cost flows directly into intelligence per dollar. NVIDIA positions Vera Rubin as optimized for this specific pattern — repeated, smaller post-training runs that never stop rather than single large training jobs.

The compute footprint grows not because individual runs are larger, but because the runs are perpetual. That changes infrastructure planning: organizations need platforms tuned for sustained, high-utilization post-training throughput, not peak single-run performance.

What This Means

These three developments point in the same direction: the cost and complexity of deploying capable AI is falling, but the operational demands are shifting rather than disappearing.

TabFM removes the per-dataset training tax for tabular ML, which is significant given how much enterprise data lives in structured tables. If zero-shot tabular inference generalizes well in production, it could reduce the data science labor currently required to maintain dozens of separate models across business units.

Inkling’s Apache 2.0 release gives enterprises a credible open-weights MoE option for on-premises or private cloud deployment — one that scores competitively on software engineering benchmarks and explicitly avoids content filtering that some organizations find problematic. The controllable thinking effort mechanism is a practical cost-management tool that closed frontier APIs don’t offer.

The NVIDIA framing around intelligence per dollar and performance per watt reflects a maturing infrastructure market. As MoE models become standard and post-training becomes continuous, data center operators face a new optimization problem: sustained throughput efficiency over time, not just peak benchmark numbers. The organizations that solve that problem at scale will have a structural cost advantage in the agentic AI era.

FAQ

What is TabFM and how does it differ from traditional tabular ML?

TabFM is a zero-shot foundation model from Google Research that generates predictions for unseen tabular datasets in a single forward pass, without retraining. Traditional tabular ML requires building a separate model per dataset, including feature engineering, hyperparameter tuning, and ongoing retraining pipelines to handle data drift.

What is Thinking Machines’ Inkling model and who can use it?

Inkling is a 975-billion-parameter multimodal mixture-of-experts model released by Thinking Machines under an Apache 2.0 open source license, meaning it is free to use and modify commercially. Weights are available on Hugging Face, and a smaller 276-billion-parameter variant called Inkling-Small is also available for cost-sensitive workloads.

Why does NVIDIA emphasize performance per watt over raw compute benchmarks?

Data centers operate under fixed power budgets, so the number of AI tokens a facility can generate per watt directly determines its revenue ceiling. Performance per watt reflects real-world throughput under actual power constraints and cannot be inflated by selective benchmark conditions, making it a more reliable planning metric than peak theoretical compute figures.