Thinking Machines Lab released Inkling, a 975-billion-parameter open-weight AI model on Wednesday that anyone can download and fine-tune, marking a deliberate shift away from the closed, metered API model that dominates the industry. Unlike flagship models from OpenAI, Anthropic, or Google, Inkling's full weights are available on Hugging Face, a popular repository for machine learning models. The company is not claiming Inkling beats every competitor on raw intelligence. Instead, it is betting that organizations will pay more for the ability to own and customize a model than to rent access to a slightly more powerful one.

What Makes Inkling Different From Claude and Other Proprietary Models?

Inkling is built as a Mixture-of-Experts transformer, a type of neural network architecture that activates only a subset of its parameters for each task, making it more efficient than traditional models. The model has 975 billion total parameters, but only 41 billion are active at any given time, allowing it to process a 1 million token context window (roughly 750,000 words) while using less computational power than its parameter count might suggest.

The company trained Inkling on 45 trillion tokens of text, images, audio, and video, giving it native multimodal reasoning from the ground up rather than bolting on vision or audio capabilities after the fact. On several benchmarks, Inkling-Small, a lighter preview version with 276 billion total parameters and 12 billion active, actually matches or beats the larger model. For example, Inkling-Small scored 83.4% on IFBench, a chat quality benchmark, compared to 79.8% for the full model.

The efficiency argument is where Inkling's design philosophy becomes clearest. By adjusting a controllable "thinking effort" setting, users can trade off speed for accuracy. Inkling matches the performance of Nemotron 3 Ultra, a competing model, on a reasoning benchmark while using roughly one-third of the tokens, or computational steps, to get there. For organizations running AI models millions of times inside longer workflows, that efficiency curve matters more than a single peak score.

How Does Inkling Perform on Real-World Tasks?

Thinking Machines trained Inkling for breadth across multiple domains rather than optimizing for a single benchmark family. The model's performance spans reasoning, coding, general agentic work, and factuality:

• Reasoning Tasks: Inkling scored 29.7% on text-only reasoning benchmarks and 46.0% when given access to tools like calculators and web search, 97.1% on the AIME 2026 mathematics competition, and 87.2% on GPQA Diamond, a difficult science benchmark.
• Coding and Software Engineering: The model achieved 77.6% on SWE-bench Verified, a benchmark for fixing real bugs in open-source code, and 63.8% on Terminal Bench 2.1, which measures agentic coding ability.
• Chat and Instruction Following: Inkling scored 79.8% on IFBench, placing it ahead of Claude Fable 5 and GPT 5.6 Sol in the company's comparison table, demonstrating strong performance on conversational tasks.
• Multimodal Capabilities: The model scored 56.6% on Audio MC, a benchmark for audio understanding, 77.2% on MMAU, and 91.4% on VoiceBench, placing it among the strongest open-weight audio models available.
• Factuality and Confidence: Inkling scored 61.1 on ForecastBench without search access, level with Gemini 3.1 Pro and ahead of Claude Opus 4.8 at 54.6, demonstrating that it learned to express appropriate confidence levels rather than always sounding certain.

The team invested significant effort into training Inkling to express the right amount of confidence, using reinforcement learning against proper scoring rules on a large corpus of real-world questions. This is a quality most AI labs treat as an afterthought, but Thinking Machines prioritized it because overconfident AI systems can mislead users and downstream applications.

How to Deploy and Customize Inkling for Your Organization

Inkling is available through multiple channels, each suited to different use cases and technical expertise levels:

• Tinker Fine-Tuning Platform: Organizations can access Inkling on Thinking Machines' Tinker platform with 64K and 256K context options at a 50% discount for a limited time. This allows teams to customize the model on proprietary data without downloading and running it themselves.
• Hugging Face Weights: Full model weights are available on Hugging Face as both the original checkpoint and as an NVFP4 checkpoint optimized for efficient inference on NVIDIA Blackwell systems, enabling organizations to run Inkling on their own infrastructure.
• Third-Party APIs: API access to Inkling is available through TogetherAI, Fireworks, Modal, Databricks, and Baseten, allowing developers to integrate the model into applications without managing infrastructure.
• Inkling Playground: A free chat interface in the Thinking Machines console lets anyone experiment with Inkling before committing to a paid run or fine-tuning project.

A real-world example demonstrates the business case for customization. Working with Bridgewater Associates, researchers used Tinker to fine-tune an open model on specialized financial data and produced a lightweight system that scored 84.7% on leading financial reasoning benchmarks, beating the most advanced proprietary alternatives while costing under 10% as much.

Why Is Thinking Machines Giving Away Its Model Weights?

The business logic behind releasing Inkling's weights for free is not charity. Once weights are public, nothing obliges a downloader to pay for inference through a metered API, so Thinking Machines' revenue model shifts from selling access to selling customization and hosting services. The company is betting that enterprises will pay more for the ability to own and fine-tune a model than to rent access to a marginally more powerful closed model.

This represents a fundamentally different bet from the metered API model that OpenAI and Anthropic depend on. Anthropic, for example, recently raised funding at a $965 billion valuation to keep up with demand for its Claude models, which are accessed through paid APIs. Thinking Machines, by contrast, is building a platform business around customization and hosting, similar to how companies like Databricks have built billion-dollar businesses around open-source Apache Spark.

The company now employs roughly 200 people, up from levels reported after departures earlier in the year. Inkling itself was built from scratch in under nine months, against the multi-year timelines its rivals run. Whether that speed holds as the model family grows remains an open question, but the rapid development suggests the company has found an efficient path to competitive models.

What Do Safety Benchmarks Reveal About Inkling's Behavior?

Safety is often treated as a secondary concern in AI development, but Thinking Machines invested in measuring and improving Inkling's behavior on adversarial and harmful inputs. The model scored 78.0% on adversarial FORTRESS, a benchmark designed to test robustness against attacks, the strongest of any open-weight model in the company's comparison table. At the same time, Inkling maintained benign scores at 95.9%, meaning it did not refuse harmless requests. The model also hit 98.6% on StrongREJECT, a benchmark for refusing harmful instructions.

The team also trained Inkling to answer directly on topics subject to censorship, and Cognition's Propaganda and Censorship Eval found strong non-compliance patterns, meaning the model did not reflexively refuse to discuss sensitive topics. This design choice reflects a deliberate philosophy that organizations downloading Inkling should have control over its behavior, rather than having safety constraints baked in by the developer.

What Technical Innovations Enable Inkling's Efficiency?

Inkling's architecture includes several technical choices that contribute to its efficiency and performance. Each Mixture-of-Experts layer carries 256 routed experts and 2 shared experts, with 6 routed experts active per token, steered by a sigmoid router with auxiliary-loss-free load balancing. This design allows the model to specialize different parts of its network for different tasks without the training instability that often plagues Mixture-of-Experts models.

Attention, the mechanism that allows the model to focus on relevant parts of its input, interleaves sliding-window and global layers at a 5 to 1 ratio with 8 key-value heads. Notably, the team found that relative positional embeddings outperformed and extrapolated better than RoPE, a widely adopted positional encoding method. Training used a hybrid optimizer, Muon for large matrix weights and Adam for the rest, on NVIDIA GB300 NVL72 systems, with reinforcement learning scaled past 30 million rollouts across two long continuous runs.

One emergent detail is worth noting. Over reinforcement learning training, the chain of thought, the step-by-step reasoning the model generates, compressed on its own, dropping articles and connectives while staying comprehensible and leaving final answers unaffected. Nothing in the reward signal targeted this behavior. Efficiency alone did the work, suggesting that the model learned to prioritize clarity and brevity as a side effect of being trained to use fewer tokens.

The release of Inkling lands in a market that has moved hard toward cost per finished task rather than peak intelligence, with rivals shipping agent products aimed squarely at enterprise buyers. Inkling-Small is still in testing, with weights due once that finishes, giving organizations another option for deploying smaller, more efficient models on resource-constrained systems.