Google Vertex AI Expands Model Garden with DeepSeek Giants: Week of 10 February 2025

Google made the week's biggest move by adding two massive 671-billion parameter DeepSeek models to Vertex AI's Model Garden, whilst simultaneously releasing Llama 3.3 70B in preview. With Qdrant delivering significant database improvements and Elastic pushing stability updates, this week showcased the industry's focus on expanding model choice and infrastructure reliability.

Google doubles down on model diversity in Vertex AI

Google's addition of DeepSeek-V3 and DeepSeek-R1 to Vertex AI Model Garden represents a significant expansion of available model architectures for enterprise developers. Both models pack 671 billion parameters, putting them in the same weight class as the largest production models available today. This move signals Google's commitment to offering model diversity rather than forcing users into a single architectural approach.

The timing is particularly interesting given the competitive pressure from providers like Anthropic and OpenAI who've been pushing their flagship models heavily. By adding DeepSeek's offerings, Google is essentially saying "we'll give you choice rather than lock you into our models". For developers, this means access to potentially different reasoning approaches and performance characteristics without leaving the Vertex AI ecosystem.

What makes this move strategically important is the 671B parameter count. These aren't lightweight models you'd run locally or use for simple tasks. They're enterprise-grade models that require serious infrastructure, which plays directly into Google's cloud computing strengths. Developers should evaluate these models against their current solutions, particularly for complex reasoning tasks where the additional parameters might provide meaningful improvements.

The simultaneous release of Llama 3.3 70B in preview further reinforces this strategy. Google is positioning Vertex AI as the Swiss Army knife of model platforms rather than trying to win with a single blade. For organisations already invested in Vertex AI infrastructure, this provides compelling reasons to consolidate model usage rather than managing multiple provider relationships.

Qdrant delivers the infrastructure improvements that actually matter

Whilst everyone focuses on flashy new models, Qdrant quietly shipped the kind of infrastructure improvements that prevent 3am wake-up calls. The headline feature is consensus compaction being enabled by default, which will accelerate peer operations and recovery times. For teams running vector databases in production, this translates to fewer headaches during scaling events and faster recovery from node failures.

The data consistency improvements tackle a more subtle but equally critical issue: preventing stale reads and writes. Anyone who's debugged inconsistent search results or wondered why their vector similarity scores seemed off will appreciate this fix. These aren't the kind of problems that make headlines, but they're the ones that make engineers question their career choices at 2am.

Qdrant's focus on operational reliability rather than feature proliferation shows maturity in the vector database space. The REST API updates and bug fixes round out a release that prioritises stability over novelty. For teams evaluating vector databases or considering migrations, Qdrant's commitment to operational excellence should factor heavily into decision-making.

Worth watching: Elastic maintains momentum with dual releases

Elastic pushed out both Elasticsearch 8.17.2 and 8.16.4 this week, continuing their aggressive maintenance schedule. These aren't feature releases but stability and bug fix updates that keep existing deployments running smoothly. The dual release approach suggests Elastic is supporting multiple version branches actively, which provides flexibility for organisations with different upgrade cadences.

For teams running Elasticsearch in production, these updates represent the unglamorous but essential work of keeping search infrastructure reliable. The 8.17.2 release particularly focuses on stability improvements that should reduce operational overhead.

Worth watching: Replicate polishes the developer experience

Replicate shipped UI improvements across deployment overview, browser navigation, and playground functionality. These changes might seem minor, but developer experience improvements often correlate with increased adoption and reduced support overhead. The accompanying terms of service updates suggest ongoing compliance maintenance.

For teams using Replicate for model deployment, these improvements should reduce friction in daily workflows. The playground enhancements particularly benefit teams doing model evaluation and testing.

Quick hits

Anthropic Economic Index launched - New research initiative tracking AI's economic impact, though practical implications remain unclear.

The week ahead: Model evaluation season begins

With Google's new DeepSeek models now available and Llama 3.3 70B in preview, expect increased focus on model evaluation and benchmarking. Teams should allocate time for testing these new options against existing solutions, particularly for complex reasoning tasks where the parameter count differences might matter.

Qdrant users should plan for updates to take advantage of the consensus compaction improvements, especially in high-throughput environments. The data consistency fixes make this a priority upgrade for production systems.

Watch for additional model releases as providers respond to Google's expanded Model Garden strategy. The competitive pressure to offer choice rather than just performance is building across the industry.