Thinking Machines develops AI models and sells Tinker, a cloud platform researchers and developers use to fine-tune models and run inference without managing the underlying GPU clusters.

Teams provide training data and write their own training loops, while Tinker executes the computations on remote machines. This gives researchers control over how a model is adapted to a domain or task, rather than limiting them to prompting a general-purpose chatbot.

On November 5, 2026, Thinking Machines launched Inkling 2 Small, a reasoning model that accepts text and images and produces text. It is available for inference and fine-tuning through Tinker, with an inference-only Playground interface; its weights are initially closed, with an Apache 2.0 release planned after external safety testing and red-teaming.

Tinker uses LoRA, a method for updating a small set of model parameters, and runs on the company's training clusters, with usage-based pricing per million tokens. Tinker exposes low-level primitives (forward_backward, optim_step, sample, save_state) and lets users download saved checkpoints for supported open-weight models, supporting a broad lineup of open-weight models including Qwen3, Qwen3-235B-A22B, Llama-3, DeepSeek-V3.1, GPT-OSS, and Kimi-K2, as well as vision-language fine-tuning. Early Tinker users include research teams at Princeton, Stanford, Berkeley, and Redwood Research.

Artificial Analysis scored Inkling 2 Small at 31 on its November 5 Intelligence Index v5.0.1, up 18 points from Inkling Small, with especially large gains on agentic tasks. Coding performance remained uneven: it beat Claude Sonnet 5.5 at medium reasoning effort on FrontierCode 2.0, but trailed it on Terminal-Bench 4.0.

The platform targets researchers, enterprises, and developers who need AI systems tailored to their specific workflows rather than generic chatbot interfaces. A pharmaceutical company could train the model on their proprietary research data and safety protocols, while a software team could customize it for their codebase and development practices.

Thinking Machines combines open-model distribution with paid cloud services for model training and inference. Developers can use Tinker to customize models and run them on the company's infrastructure, while the planned release of Inkling 2 Small's weights would expand the options for running that model elsewhere.

Tinker charges for usage, so customer spending depends on the work performed during training and the tokens processed during inference. Making a base model available to developers can draw them into the training platform, but downloaded weights do not by themselves generate hosting revenue.

Inkling 2 Small's stronger benchmark scores also come with heavier inference workloads. Artificial Analysis measured approximately 150,000 output tokens and $1.18 per Intelligence Index task, versus 37,000 tokens and $0.20 for Inkling Small; these are benchmark costs, rather than a typical customer bill or a measure of Thinking Machines' margins.

The company must pay for the GPU capacity and technical staff supporting training and inference. Its economics depend on charging enough for this work to cover those costs, even when customers can move open-weight models to another provider.

Developing its own models alongside Tinker gives Thinking Machines a way to improve both the models and the tools used to customize them. Whether that translates into durable pricing power depends on researchers continuing to choose its hosted services over competing training platforms and self-managed infrastructure.

OpenAI, Google DeepMind, and Anthropic dominate the frontier AI landscape through massive scale and vertical integration. OpenAI's GPT-4 family benefits from Microsoft's Azure distribution and aggressive pricing at $5 per million tokens, while Google embeds Gemini directly into Workspace and Android.

Anthropic's Claude models focus on safety and compliance, particularly for financial services, backed by over $7 billion from Amazon and Google. These players can subsidize model development through their broader ecosystems, making it difficult for standalone companies to compete on price or distribution.

Mistral AI represents the European approach to AI sovereignty, raising over €1 billion while releasing both open-weight and commercial models. Their focus on smaller, efficient models that rival GPT-4 performance creates pricing pressure for larger players.

Meta's Llama family and Chinese companies like DeepSeek push open-source alternatives that enterprises can deploy without vendor lock-in. These competitors threaten Thinking Machines' positioning by offering similar customization capabilities without the premium pricing of frontier models.

AWS Bedrock, Microsoft Azure AI, and Google Vertex AI provide model-agnostic platforms that let enterprises access multiple foundation models through unified APIs. Hugging Face has built the dominant open-source ecosystem for model deployment and fine-tuning, while NVIDIA's AI foundry services target enterprises wanting custom models.

These platforms compete directly with Thinking Machines' customization layer, offering similar fine-tuning and deployment capabilities across multiple model providers rather than being locked into a single foundation model.

Thinking Machines plans to expand beyond their core foundation model into specialized AI systems for scientific research, drug discovery, and advanced engineering applications. The company is developing domain-specific models trained on scientific literature, laboratory data, and research methodologies that could command premium pricing from pharmaceutical companies, materials science labs, and academic institutions.

Their roadmap includes AI systems for protein folding, molecular design, and climate modeling that would tap into the growing market for AI-powered scientific discovery.

The platform's customization capabilities enable expansion from AI research labs to mainstream enterprise R&D departments across industries. Fortune 500 companies in automotive, aerospace, and manufacturing represent a massive untapped market for AI systems that can be trained on proprietary engineering data and safety protocols.

The company's open-source components also create a pathway to millions of individual developers and smaller companies that currently use general-purpose models but need more specialized capabilities for their specific use cases.

Thinking Machines' public benefit corporation structure and European investment provide entry points into international markets with strong AI sovereignty requirements. The EU's AI Act and similar regulations in other countries create demand for transparent, auditable AI systems that can be customized for local compliance requirements.

The company's distributed team and open-source approach enable rapid localization for non-English markets without requiring full regional offices, potentially capturing market share in regions where US-based AI companies face regulatory or political barriers.

Compute access: Thinking Machines' success depends entirely on securing massive GPU capacity in a market dominated by cloud providers who also compete in AI. NVIDIA's H100 and H200 chips remain scarce, and hyperscalers like Microsoft, Google, and Amazon prioritize their own AI initiatives over third-party customers.

Talent attrition: Thinking Machines faces an accelerating exodus of founding members, with departures flowing to both OpenAI and Meta — two of the best-resourced talent competitors in the industry. Each founding-level departure thins the technical leadership bench that underpins the company's core differentiation.

Open source commoditization: Thinking Machines' strategy of releasing open-source components to drive adoption could backfire if competitors use their own tools and research to build competing platforms, as Meta's approach with Llama has demonstrated by rapidly commoditizing entire market segments. This dynamic risks eliminating the pricing power that justifies Thinking Machines' massive valuation before the company establishes a sustainable competitive moat.