CtrlK
IBM watsonx is an artificial intelligence and data platform developed by IBM for enterprise use. Announced on May 9, 2023, at IBM's annual Think conference by CEO Arvind Krishna, watsonx represents IBM's current generation of AI products, succeeding the earlier Watson platform that gained fame through its 2011 Jeopardy! victory. The platform consists of three integrated components: watsonx.ai for model training and deployment, watsonx.data for data management, and watsonx.governance for AI compliance and risk management. IBM's proprietary Granite family of large language models, released under the permissive Apache 2.0 license, serves as the core model offering within the watsonx ecosystem.
IBM has been involved in artificial intelligence research for over seven decades, making it one of the longest-running participants in the field. The company's AI efforts date back to the 1950s, when IBM researcher Arthur Samuel developed a checkers-playing program that could learn from its own experience, coining the term "machine learning" in 1959.
IBM achieved its first widely recognized AI milestone with Deep Blue, a chess-playing supercomputer that defeated reigning world champion Garry Kasparov in a six-game match in May 1997. Deep Blue was capable of evaluating 200 million chess positions per second, achieving a processing speed of 11.38 billion floating-point operations per second. The system won two games and drew three in the rematch, after losing to Kasparov 4-2 in their first encounter in 1996. The victory is broadly considered a landmark moment in AI history, demonstrating that machines could surpass human performance in complex strategic tasks.
In February 2011, IBM's Watson question-answering system competed on the television quiz show Jeopardy! against the show's two greatest all-time champions, Ken Jennings and Brad Rutter. Watson won with $77,147 in prize money (donated to charity), compared to Jennings' $24,000 and Rutter's $21,600. The system was developed by an IBM research team led by principal investigator David Ferrucci, who had pitched the idea of building a Jeopardy!-playing computer back in 2006.
The Watson system that competed on Jeopardy! was a room-sized computer consisting of 10 racks holding 90 servers with a total of 2,880 processor cores. It ingested massive amounts of unstructured text from Wikipedia, encyclopedias, dictionaries, novels, plays, and other sources from Project Gutenberg. Unlike a traditional search engine, Watson could understand natural language questions and generate precise answers without an internet connection. The victory represented a significant advance in natural language processing and demonstrated the potential for intelligent machines to analyze unstructured data at scale.
Following the Jeopardy! triumph, IBM invested heavily in commercializing Watson as an enterprise AI platform. The company launched Watson Health in April 2015 with ambitious goals for the healthcare market, particularly in oncology. IBM spent approximately $5 billion acquiring health data companies, including Truven Health Analytics, Phytel, Explorys, and Merge Healthcare, to build Watson's healthcare data capabilities.
However, Watson Health struggled to deliver on its promises. In 2015, MD Anderson Cancer Center ended its collaboration with IBM Watson after spending $62 million over two years. By 2018, more than a dozen IBM partners and clients had stopped or scaled back their oncology projects with Watson. In April 2019, IBM halted development and sales of its Watson AI drug discovery tools due to disappointing results. Critics pointed to the gap between Watson's marketing claims and its actual clinical utility, as the system often produced unreliable or unsafe treatment recommendations.
On January 21, 2022, IBM announced the sale of Watson Health's core data and analytics assets to private equity firm Francisco Partners for more than $1 billion. The acquired assets included Health Insights, MarketScan, Clinical Development, Social Program Management, Micromedex, and imaging software offerings. Francisco Partners subsequently rebranded these assets as Merative. The sale price represented a fraction of the roughly $5 billion IBM had invested in healthcare data acquisitions, marking a significant financial loss and the end of IBM's attempt to build a healthcare-focused AI business under the Watson brand.
IBM unveiled watsonx at its Think 2023 conference on May 9, 2023, positioning the platform as a fresh start for its enterprise AI strategy in the era of generative AI. According to Arvind Krishna, "Foundation models make deploying AI significantly more scalable, affordable, and efficient. We built IBM watsonx for the needs of enterprises, so that clients can be more than just users, they can become AI advantaged."
The platform became generally available in phases throughout 2023, with watsonx.ai and watsonx.data launching in July 2023 and watsonx.governance following in early December 2023.
watsonx.ai is IBM's AI studio for building, training, validating, tuning, and deploying both traditional machine learning models and generative AI capabilities powered by foundation models. The studio provides:
IBM has steadily expanded the range of third-party models available on watsonx.ai. In February 2024, IBM added the Mixtral-8x7B model from Mistral AI. Meta's Llama 3 became available in April 2024, and Llama 3.2 models (including multimodal 11B and 90B variants) were added later that year. Mistral Large 2 was introduced to the platform in July 2024.
watsonx.data is a data lakehouse built on open architecture designed to manage and optimize data for AI workloads. It supports open data formats such as Apache Iceberg and Apache Parquet, enabling organizations to access, share, and govern data across multiple environments. IBM claims organizations can reduce their data warehouse costs by up to 50 percent through workload optimization with watsonx.data. The platform supports multiple query engines and integrates with existing data infrastructure, allowing enterprises to run analytics and AI workloads without needing to move or copy all their data to a single location.
watsonx.governance is IBM's AI governance toolkit, providing automated workflows and tools for managing AI models throughout their lifecycle. It enables organizations to:
The governance component is particularly relevant for enterprises operating in regulated industries such as financial services, healthcare, and government, where AI transparency and accountability are required by law or internal policy.
Granite is IBM's family of proprietary AI models, first announced on September 7, 2023, with the initial models (Granite.13b.instruct and Granite.13b.chat, each with 13 billion parameters) becoming generally available on September 28, 2023. All Granite models are released under the Apache 2.0 open-source license, which permits unrestricted commercial use, modification, and redistribution. This licensing approach distinguishes IBM from many competitors who release models under more restrictive terms.
IBM trains Granite models on carefully curated, license-permissible data, following IBM's AI Ethics principles and guided by IBM's Corporate Legal team. This approach is designed to reduce intellectual property risks for enterprise customers.
On May 6, 2024, IBM released the Granite Code model family, a set of decoder-only models trained for code generation tasks across 116 programming languages. The models are available in sizes of 3B, 8B, 20B, and 34B parameters, in both base and instruct variants. Granite Code models are designed for code generation, bug fixing, code explanation, documentation, application modernization, and repository maintenance. IBM also open-sourced the data-prep-kit framework and pipelines used to prepare the training data.
On October 21, 2024, IBM introduced Granite 3.0, a significantly expanded model family. The release includes:
| Model | Type | Parameters | Active Parameters | Context Length | Training Data |
|---|---|---|---|---|---|
| Granite 3.0 8B Instruct/Base | Dense LLM | 8B | 8B | 4K tokens | 12T tokens |
| Granite 3.0 2B Instruct/Base | Dense LLM | 2B | 2B | 4K tokens | 12T tokens |
| Granite 3.0 3B-A800M Instruct | MoE | 3B | 800M | 4K tokens | 10T tokens |
| Granite 3.0 1B-A400M Instruct | MoE | 1B | 400M | 4K tokens | 10T tokens |
| Granite Guardian 3.0 8B | Safety | 8B | 8B | 4K tokens | Annotated safety data |
| Granite Guardian 3.0 2B | Safety | 2B | 2B | 4K tokens | Annotated safety data |
The dense 8B model rivaled Meta's Llama 3.1 8B Instruct across both OpenLLM Leaderboard v1 and v2 benchmarks. The Mixture of Experts (MoE) models use significantly fewer active parameters at inference time, making them suitable for on-device deployment and low-latency scenarios.
Released on December 18, 2024, Granite 3.1 expanded context windows across the entire model family to 128K tokens. The release maintained the same model sizes as 3.0 (dense 8B and 2B, MoE 3B-A800M and 1B-A400M) while adding longer context support and improved performance.
| Model | Type | Parameters | Context Length | Key Improvement |
|---|---|---|---|---|
| Granite 3.1 8B Instruct | Dense LLM | 8B | 128K tokens | Extended context |
| Granite 3.1 2B Instruct | Dense LLM | 2B | 128K tokens | Extended context |
| Granite 3.1 3B-A800M Instruct | MoE | 3B | 128K tokens | Extended context |
| Granite 3.1 1B-A400M Instruct | MoE | 1B | 128K tokens | Extended context |
| Granite Guardian 3.1 8B | Safety | 8B | 128K tokens | Extended context |
| Granite Guardian 3.1 2B | Safety | 2B | 128K tokens | Extended context |
Announced on February 26, 2025, Granite 3.2 introduced two significant new capabilities: reasoning and multimodal vision.
The text-only instruct models (8B and 2B) gained reasoning capabilities, allowing them to produce chain-of-thought outputs before generating final answers. IBM also released Granite Vision 3.2 2B, a lightweight multimodal model specifically designed for document understanding tasks. The vision model was trained using IBM's open-source Docling toolkit to process 85 million PDFs and generate 26 million synthetic question-answer pairs. Despite having only 2 billion parameters, IBM reported that the vision model matched or exceeded the performance of models five times its size (such as Llama 3.2 11B and Pixtral 12B) on enterprise benchmarks including DocVQA, ChartQA, AI2D, and OCRBench.
Released on April 16, 2025, Granite 3.3 introduced Granite Speech 3.3 8B, a speech-to-text model with translation capabilities. The text models (Granite 3.3 8B Instruct) received improved reasoning and fill-in-the-middle (FIM) capabilities, with significant performance gains on AlpacaEval-2.0 and Arena-Hard benchmarks. The models support structured reasoning through dedicated think and response tags, and handle 12 languages including English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.
Granite 4.0, released on October 2, 2025, introduced a hybrid Mamba-2/Transformer architecture that combines standard transformer-style attention layers with Mamba-2 layers. This architectural change makes the models significantly faster and more memory-efficient than comparably-sized pure-transformer models.
| Model | Parameters | Architecture | Key Feature |
|---|---|---|---|
| Granite 4.0 H-Small | ~14B | Hybrid Mamba-2 | Flagship enterprise model |
| Granite 4.0 H-Tiny | 7B | Hybrid Mamba-2 | Balanced performance |
| Granite 4.0 H-Micro | 3B | Hybrid Mamba-2 | Efficient deployment |
| Granite 4.0 Nano (1.5B-350M) | 350M-1.5B | Hybrid Mamba-2 | On-device / browser |
The Nano models, released on October 29, 2025, are small enough to run on laptops and directly in web browsers. Enterprise partners including EY and Lockheed Martin received early access for testing.
Granite Guardian is a specialized sub-family of Granite models designed for AI safety and risk detection. These models serve as guardrails for LLM applications, detecting risks in both user prompts and model responses.
Granite Guardian assesses multiple risk dimensions:
In testing across 19 safety and RAG benchmarks, Granite Guardian 3.0 8B demonstrated higher overall accuracy on harm detection than all three generations of Meta's Llama Guard models. For hallucination detection, it showed performance on par with specialized models like WeCheck and MiniCheck.
Granite Guardian models are trained on a combination of human-annotated and synthetic data, with prompt-response pairs annotated for different risk dimensions by a socioeconomically diverse group of annotators at DataForce.
Beyond language and code models, IBM has expanded the Granite family into specialized domains.
Granite Embedding models generate vector representations of text inputs for semantic search and retrieval-augmented generation (RAG) applications. The R2 release introduced models based on the ModernBERT architecture: granite-embedding-english-r2 (149M parameters, 768-dimensional embeddings) and granite-embedding-small-english-r2 (47M parameters, 384-dimensional embeddings). These models were trained on 2 trillion tokens from high-quality web-based corpus and code data.
Granite Time Series models are pre-trained on time series data for forecasting tasks. IBM reported that updated versions, trained on three times more data than earlier releases, outperformed models ten times their size from Google, Alibaba, and others on major time series benchmarks.
On December 5, 2023, IBM and Meta co-founded the AI Alliance, a global coalition dedicated to promoting open, safe, and responsible AI development. The alliance launched with more than 50 founding members and has since grown to over 180 organizations as of late 2024.
Founding members include CERN, Dell Technologies, Hugging Face, Intel, Oracle, Sony Group, AMD, and a broad range of universities worldwide. Notable companies that did not join include AWS, Google, Microsoft, NVIDIA, and OpenAI.
The AI Alliance focuses on several key areas:
The Safety and Trust working group, launched in 2024, grew to more than 230 individual participants from over 40 organizations. The alliance also initiated a collaboration with IBM, Red Hat, Mass Open Cloud Consortium, and the National Science Foundation to create an open AI cloud environment for the research community.
The formation of the AI Alliance reflects IBM's broader strategic position favoring open-source AI development, in contrast to the more closed approaches taken by companies such as OpenAI and Google.
IBM's AI strategy centers on serving regulated industries where governance, data privacy, and compliance are critical requirements. Rather than competing directly with OpenAI or Google on consumer-facing AI products, IBM targets enterprise customers in financial services, healthcare, government, telecommunications, and manufacturing.
IBM positions watsonx through several enterprise-focused advantages:
IBM's 2019 acquisition of Red Hat for $34 billion plays a significant role in its AI strategy. Red Hat Enterprise Linux AI (RHEL AI) packages Granite models for deployment on Red Hat's infrastructure, and Red Hat OpenShift AI provides a platform for running AI workloads in hybrid cloud and on-premises environments. This integration gives IBM a distribution channel for Granite models that most AI competitors lack.
IBM reported total annual revenue of $62.8 billion for fiscal year 2024, a 1.4% increase from the prior year. The company's software segment, which includes watsonx, grew 8.3% year-over-year and represented approximately 45% of total revenue by the end of 2024. Consulting revenue was $20.7 billion, while infrastructure revenue declined 3% at constant currency.
IBM's generative AI-related book of business has grown rapidly. The company reported cumulative generative AI bookings exceeding $6 billion as of Q1 2025, growing from a base of $5 billion just one quarter earlier. Approximately 80% of these bookings came from the Consulting segment, with the remainder from Software. IBM highlighted over 200 consulting projects using AI agents at scale and over 1,000 client engagements.
By early 2026, IBM reported generative AI bookings exceeding $12 billion cumulatively. Analysts have projected that IBM's AI revenue could reach $15 billion by 2027.
IBM watsonx competes with several major enterprise AI platforms:
| Platform | Provider | Key Strength | Primary Audience |
|---|---|---|---|
| watsonx | IBM | Governance, hybrid deployment, open-source models | Regulated enterprises |
| Azure AI | Microsoft | OpenAI integration, Copilot ecosystem | Broad enterprise |
| Vertex AI | Gemini models, data analytics integration | Cloud-native enterprises | |
| Amazon Bedrock | AWS | Multi-model marketplace, AWS ecosystem | AWS customers |
| Nvidia AI Enterprise | NVIDIA | GPU infrastructure, inference optimization | AI-intensive workloads |
IBM's competitive advantage lies in serving organizations with strict regulatory requirements, complex hybrid infrastructure, and a need for transparent, governable AI. The company's long history in enterprise technology, combined with its consulting capabilities, gives it credibility with large organizations that may be cautious about adopting AI from newer, consumer-focused companies.
However, IBM faces challenges. Its overall revenue growth lags behind cloud-native competitors, and the company's AI brand was damaged by the Watson Health experience. Microsoft's partnership with OpenAI and Google's vertically integrated Gemini ecosystem command larger shares of the generative AI market. IBM's smaller model sizes (topping out at 8B-14B parameters for Granite) also mean it relies on third-party models for the most demanding tasks.
| Date | Event |
|---|---|
| May 1997 | Deep Blue defeats Garry Kasparov in chess |
| February 2011 | Watson defeats Ken Jennings and Brad Rutter on Jeopardy! |
| April 2015 | IBM launches Watson Health |
| July 2019 | IBM acquires Red Hat for $34 billion |
| January 2022 | IBM sells Watson Health assets to Francisco Partners for over $1 billion |
| May 9, 2023 | IBM unveils watsonx platform at Think 2023 |
| July 2023 | watsonx.ai and watsonx.data become generally available |
| September 7, 2023 | IBM announces first Granite foundation models |
| September 28, 2023 | Granite 13B models become generally available |
| December 2023 | watsonx.governance launches; IBM and Meta co-found AI Alliance |
| May 2024 | Granite Code models released under Apache 2.0 |
| October 21, 2024 | Granite 3.0 released (8B, 2B dense; MoE; Guardian models) |
| December 18, 2024 | Granite 3.1 released with 128K context windows |
| February 26, 2025 | Granite 3.2 released with reasoning and vision capabilities |
| April 16, 2025 | Granite 3.3 released with speech model and improved reasoning |
| October 2, 2025 | Granite 4.0 released with hybrid Mamba-2 architecture |