The “first AI product is in the works” headline is now outdated. Mira Murati’s Thinking Machines Lab first launched Tinker, a managed platform for training and fine-tuning AI models, in October 2025. It later released Inkling, its first internally trained open-weights model, in July 2026.
That distinction matters for anyone evaluating the company as an AI investment story, technology supplier, or potential business customer: Tinker is the commercial platform, while Inkling is the company’s model. Neither is a consumer chatbot, and Inkling’s “open” status does not make it inexpensive to operate.
As an Amazon Associate I earn from qualifying purchases.
The short answer
| Product | What it is | Launch | Likely user |
|---|---|---|---|
| Tinker | Managed API and infrastructure for model training and fine-tuning | October 1, 2025; generally available December 12, 2025 | Researchers, developers and AI startups |
| Inkling | Thinking Machines Lab’s first internally trained open-weights multimodal model | July 15, 2026 | Organizations and technical teams building customized AI systems |
| Inkling-Small | A smaller, lower-cost model with multimodal input and variable thinking effort | July 30, 2026 | Teams prioritizing cost, latency and customization |
So, if “first product” means the company’s first commercial platform, the answer is Tinker. If it means the first model trained and released by Thinking Machines Lab, the answer is Inkling.
What was announced in July 2025?
On July 16, 2025, Mira Murati said Thinking Machines Lab expected to release its first product within the next couple of months. The announcement described a significant open-source component, a focus on multimodal AI, and usefulness for researchers and startups developing custom models.
#1 Best Overall
At that point, the company had not disclosed the product’s name, user interface, pricing, model size, launch date or technical specifications. It would therefore be inaccurate to treat the original announcement as a description of Inkling specifically. The later product strategy became clearer only after Tinker and Inkling were announced.
The company had also attracted substantial attention before shipping a public product. Contemporary reports put its 2025 seed financing above $2 billion and its reported valuation at $12 billion, with investors including Andreessen Horowitz, NVIDIA, Accel, ServiceNow, Cisco and AMD. Those are reported financing and valuation figures—not revenue, profit or evidence that the company had achieved product-market fit. See the original product-announcement coverage for that context.
What is Tinker?
Tinker is a training API for researchers and developers. Instead of requiring customers to build and operate a distributed GPU cluster, Thinking Machines Lab manages the underlying compute and infrastructure while users control important parts of the training process through code.
Free tools Windows power users keep installed
One-click scans. No signup required.
Its documented core operations include:
forward_backward: runs forward and backward passes and accumulates gradients.optim_step: updates model weights.sample: generates tokens for evaluation, interaction or reinforcement-learning actions.save_state: saves training progress for later continuation.
This makes Tinker materially different from a normal chatbot API or a no-code fine-tuning wizard. It is intended for specialized agents, search and data-processing systems, forecasting, continual-learning experiments, reinforcement learning, post-training and academic or industrial research.
Tinker became generally available on December 12, 2025, when the company said the waitlist had ended. Its documentation includes a Python SDK, training and sampling clients, checkpoint management, billing commands, LoRA workflows, several reinforcement-learning loss functions, and OpenAI-compatible and Anthropic-compatible interfaces. Availability and supported models can change, so users should check the current documentation before committing to a project.
What is Inkling?
Inkling is described by Thinking Machines Lab as a general-purpose, open-weights Mixture-of-Experts transformer. According to the company’s announcement and model card, it has:
Rank #2
- 975 billion total parameters.
- Approximately 41 billion active parameters per task.
- Training on 45 trillion tokens across text, images, audio and video.
- Native reasoning over text, images and audio.
- Text generation as its output modality.
- Weights available under the Apache 2.0 license.
- Fine-tuning available through Tinker.
“Multimodal” here means that Inkling can accept multiple types of input. It does not mean the model generates images, audio or video. The published model information describes text output.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Thinking Machines Lab also says Inkling is not the strongest overall model available, whether among open or closed models. Its proposed advantage is customization: developers can adapt it to their own data, workflows and preferred reasoning behavior rather than simply asking a general model for an answer.
Open weights are not the same as fully open-source AI
Inkling’s Apache 2.0 license and downloadable weights are significant, but “open weights” should not automatically be treated as “fully reproducible open-source training.” The available weights do not necessarily include the complete training dataset, filtering process, data mixture, infrastructure or training code needed to recreate the model from scratch.
There is also a practical difference between owning a copy of the weights and having a portable operating environment. A business can download Inkling, but reproducing the managed training experience locally requires a large amount of hardware and engineering expertise.
Can you run Inkling locally?
For most individuals and small companies, not realistically. The model card says the BF16 checkpoint requires at least 2 TB of aggregated VRAM. Examples listed by the company include eight NVIDIA B300 GPUs or 16 NVIDIA H200 GPUs.
The NVFP4 quantized checkpoint requires at least 600 GB of aggregated VRAM, with listed configurations including four NVIDIA B300 GPUs in W4A4 mode or eight NVIDIA H200 GPUs in W4A16 mode.
Those figures are minimum aggregate memory requirements, not a guarantee of convenient deployment speed, low operating cost or simple setup. They are far beyond a typical gaming PC, developer laptop or ordinary single-server workstation. The model card identifies deployment frameworks including SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face, but using those frameworks still requires suitable hardware and deployment expertise.
For many prospective users, the practical options are Tinker, a third-party inference provider, or rented cloud infrastructure. The Hugging Face weights make the model more portable, but they do not eliminate the cost of running it.
What does Tinker cost?
Tinker uses usage-based pricing rather than a simple consumer subscription. Charges can include training tokens, sampling, prefill, cached prefill and checkpoint storage. The documentation lists checkpoint storage at $0.10 per GB per month and says cached prefill receives an 80% discount.
The following figures were displayed on August 18, 2026 and included a limited-time 50% discount. They should not be treated as permanent rates:
| Model | Context | Prefill | Cached prefill | Sample | Train |
|---|---|---|---|---|---|
| Inkling | 64K | $1.87/M | $0.374/M | $4.68/M | $5.61/M |
| Inkling | 256K | $3.74/M | $0.748/M | $9.36/M | $11.23/M |
| Inkling-Small | 64K | $0.58/M | $0.116/M | $1.44/M | $1.73/M |
| Inkling-Small | 256K | $1.16/M | $0.232/M | $2.89/M | $3.47/M |
“M” means one million tokens. A long training run can therefore cost substantially more than a few ordinary API calls, particularly when it involves repeated sampling, large datasets and checkpoint storage. Before using proprietary data, organizations should also review retention, privacy, security and contractual terms.
Tinker’s documentation lists serverless inference for Inkling and Inkling-Small as beta. It says intensive production use is not recommended until that service leaves beta, and some production users may need to join a waitlist. General availability of the platform should not be confused with a blanket guarantee of production readiness for every workload.
What is Inkling-Small?
Inkling-Small was announced on July 30, 2026 as a smaller and lower-cost option. Thinking Machines Lab says it supports native reasoning over audio and images, variable thinking effort, a context window of up to one million tokens, full-weight availability and fine-tuning through Tinker. Text, image and audio access is also available through the Tinker Playground.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →The company reported output pricing of $4.05 per million tokens for Inkling and $1.20 per million tokens for Inkling-Small, while noting that Tinker discounts were available for a limited time. “Small” should not be read as “runs on any consumer computer”; deployment requirements and performance must be assessed separately from the model’s name.
Why the one-million-token claim needs qualification
The company has advertised context windows of up to one million tokens for Inkling and Inkling-Small. However, the current Tinker listings referenced in the pricing documentation show 64K and 256K context options for the models.
That means readers should distinguish between the model’s stated maximum context and the configuration currently available through a particular service. A one-million-token capability may not be available through every provider, checkpoint or API configuration.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Who is Tinker for?
Academic and industrial researchers
Tinker may be useful when a research project needs control over training, sampling, checkpoints or reinforcement-learning experiments without requiring the team to operate a complete multi-GPU cluster.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →AI startups
Startups building domain-specific agents or data-processing systems may value managed infrastructure and the ability to customize open models. The trade-off is variable spending and dependence on a supplier’s quotas, supported models and pricing.
Best Value
Enterprise machine-learning teams
Enterprises should evaluate Tinker alongside security, governance, data-handling and procurement requirements. A managed platform may reduce infrastructure work, but it may not satisfy teams that require fully self-hosted training or dedicated infrastructure from the outset.
Independent developers
Technically sophisticated developers can use the SDK and cookbook examples to experiment with fine-tuning and post-training. It is not a no-code product, and token and storage costs need to be monitored.
Ordinary consumers
Tinker and Inkling are not primarily consumer products. Someone who simply wants a general chatbot is unlikely to benefit from training APIs, checkpoint management or multi-GPU model deployment.
How does this compare with ordinary AI APIs?
Ordinary AI APIs generally emphasize sending a prompt and receiving an answer. Tinker emphasizes changing the model’s behavior through training, sampling and post-training workflows.
That creates a different value proposition. Tinker may be attractive when a company has proprietary examples, preference data, evaluation traces or a specialized workflow that prompt engineering cannot adequately capture. It may be excessive when the need is occasional text generation, summarization or question answering.
Alternatives include Hugging Face for model access and hosting, NVIDIA DGX Cloud for managed GPU infrastructure, AWS SageMaker and Google Cloud Vertex AI for broader enterprise platforms, and Together AI for hosted open-model inference and fine-tuning. The right choice depends on whether the priority is low-level research control, cloud integration, dedicated infrastructure, model breadth or production inference.
Financial and business risks to consider
- Unpredictable usage costs: Training, sampling and storage charges can grow with experimentation and traffic.
- Vendor dependence: Users rely on Thinking Machines Lab for capacity, service availability, supported models and pricing.
- Model portability limits: Downloadable weights do not make local operation inexpensive.
- Production uncertainty: Serverless inference remains labeled beta in the cited documentation.
- Benchmark uncertainty: Company-reported results should be compared with independently reproduced tests and the relevant checkpoint and evaluation method.
- Long-term support: The documentation records model retirements, so a project should account for migration risk.
- Data governance: Proprietary training data and production traces require a review of privacy, security and retention terms.
The company’s March 2026 announcement of a long-term, gigawatt-scale strategic partnership with NVIDIA also signals an ambition to build at significant infrastructure scale. That is strategically relevant, but it does not by itself establish customer traction, profitability or investment value.
Recommended Free Tools
Thinking Machines Lab product timeline
- July 16, 2025: Murati says the first product is expected in the coming months and will include a substantial open-source component.
- October 1, 2025: Tinker is announced.
- December 12, 2025: Tinker reaches general availability and its waitlist ends.
- March 10, 2026: The company announces a long-term strategic partnership with NVIDIA.
- July 15, 2026: Inkling is released as the company’s first open-weights model.
- July 30, 2026: Inkling-Small is announced.
Bottom line
Thinking Machines Lab’s first product is no longer merely “in the works.” Tinker was the company’s first commercial product, and Inkling was its first internally trained open-weights model. Together they reveal a strategy built around managed AI-training infrastructure plus models that organizations can customize.
For researchers and AI startups, that combination may be valuable because it reduces the need to operate distributed GPU infrastructure while preserving more training control than a standard chatbot API. For consumers and small teams seeking a cheap, ready-to-use assistant, it is the wrong category of product. Inkling’s open weights are technically notable, but its published hardware requirements and Tinker’s usage billing make total cost—not just license terms—the key financial question.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
