Mira Murati, who departed OpenAI as Chief Technology Officer, has released her startup's second-generation enterprise AI model. On November 5, 2024, Thinking Machines Lab introduced Inkling 2 Small, a 594-billion-parameter AI model designed for enterprise workflows including software development, document analysis, and visual reasoning.

The model's release positions Murati as a potential infrastructure provider for organizations seeking alternatives to closed commercial platforms, with an Apache 2.0 open-source license planned pending safety evaluations. However, the weights remain closed at launch, with the company targeting their release for the week following the announcement.

Near-Frontier Performance with Notable Gaps

Inkling 2 Small achieved third place on EnterpriseBench-CoreCraft with a 71.8% score, trailing only Anthropic's Claude Fable 5.1 (77.4%) and Claude Opus 5.5 (74.4%). This benchmark measures agentic AI performance across a fictional PC parts supplier scenario, including simulated dissatisfied customers. The model utilizes a sparse architecture that activates 17 billion of its 594 billion total parameters per token—a significant upgrade from the original Inkling Small released in July 2024, which featured 276 billion parameters with 12 billion active per token.

On the Artificial Analysis Intelligence Index, Inkling 2 Small scored 31 points, which Thinking Machines claims will be the highest among non-Chinese open models when weights are released, exceeding Mistral Large 4's score of 26. The model also achieved 46.8% on EnterpriseOps-Gym benchmarks across Customer Service, HR, and IT verticals, outperforming both GLM-5.3 and Claude Sonnet 5.

However, significant weaknesses emerged in specialized document processing tasks. The model scored just 25.2% on GDP.pdf document processing, 14.3% on the DAYJOB financial test, and 14.5% pass@1 on Apex-Accounting v1.1. Thinking Machines notes that demonstrations "do not establish reliability on an organization's production projects" and that the 25.2% GDP.pdf score is "a relative strength in this comparison, not evidence that document processing can proceed without validation."

Additionally, the model generated 154 million output tokens during Intelligence Index evaluation—72% more than the median 89.8 million among comparable models, raising concerns about potential verbosity in production settings requiring concise outputs.

Infrastructure and Deployment

Thinking Machines will offer Inkling 2 Small through its Tinker platform at $0.30 per million input tokens, $0.03 for cached input, and $1.00 for output tokens. The serverless beta is currently available, with a production waitlist active.

For organizations considering self-hosting, NVIDIA's documentation for its comparable Nemotron 3 Ultra model lists configurations including four B200 or B300 GPUs, or eight H100s. While these requirements apply to Nemotron rather than Inkling directly, they illustrate the hardware class involved in serving large sparse models of this scale.

The model's training data included software repositories, terminal environments, document production, research tasks, visual tasks, and simulated user interactions. Inkling 2 Small is multimodal, processing text and images but not audio.

Customization as Core Value Proposition

Thinking Machines built Inkling 2 Small to improve through additional training, rather than relying exclusively on its initial capabilities. In one financial-document evaluation, the company reports the new model reached the older Inkling Small's final performance after 75 training steps and then continued improving.

The company also demonstrated the model directing a training experiment, where Inkling used Tinker to adapt Qwen3.5-4B for forecasting, building an evaluator, experimenting with training methods, and debugging the process. Thinking Machines reports that the forecasting score more than doubled after 4.5 hours and $426 in Tinker spending.

The planned Apache 2.0 license enables customization without vendor lock-in—a significant advantage for organizations with specialized protocols or repeatable tasks with clear scoring criteria. However, organizations should weigh the capital-intensive infrastructure requirements against API-based services, particularly given Claude's superior EnterpriseBench performance.

Enterprise Implications

For enterprise AI teams, Murati's entry creates a potential alternative for self-hosted AI systems with complete data sovereignty. The Apache 2.0 license, once weights are released, offers a path from hosted experiments to customized models that organizations can operate themselves.

However, the documented weaknesses on financial document processing, high token generation, and the qualification that benchmark results do not establish production reliability suggest Inkling 2 Small requires careful evaluation for specific use cases. The company notes that the useful accounting unit is a completed task that meets acceptance criteria rather than raw per-token pricing.

Murati's strategic pivot from OpenAI to enterprise-focused open models signals the maturation of enterprise AI from experimental tools to operational infrastructure decisions, where data sovereignty and customization increasingly compete with raw performance metrics.

Sources

  1. Thinking Machines launches Inkling 2 Small with near-frontier performance on enterprise benchmarks — VentureBeat