Murati's Thinking Machines ships Inkling-Small at about one-quarter of Inkling's size

Why it matters

Murati is turning an expensive frontier model into a cheaper open-weight family that can distribute Thinking Machines' technology and pull developers into Tinker.

Mira Murati's Thinking Machines Lab released Inkling-Small on July 30, giving the AI lab a second open-weight model and a cheaper foundation for the fine-tuning platform at the center of Murati's AI strategy.

The release, reported by VentureBeat and detailed in Thinking Machines' launch post, came 15 days after Murati introduced the original Inkling. Inkling-Small has 276 billion total parameters and activates 12 billion for each token, compared with Inkling's 975 billion total and 41 billion active parameters. Thinking Machines says the smaller model reaches comparable aggregate performance with substantially less compute.

Murati, Thinking Machines' cofounder and CEO, is a Dartmouth-trained engineer who worked at Tesla before joining OpenAI, where she became chief technology officer and briefly served as interim CEO during the company's November 2023 leadership crisis. She formed Thinking Machines around a thesis that advanced AI should be easier for people and organizations to shape. The stated mission is to build systems that extend "human will and judgment."

That thesis was still largely a promise when RuntimeWire covered Murati's interface bet in June. Thinking Machines had previewed continuous interaction across audio, text and video, while leaving the timing of its models unstated. The July releases put downloadable weights underneath that vision and connect them to Tinker, Thinking Machines' service for fine-tuning models without managing the underlying GPU infrastructure.

Efficiency is the product decision

Inkling-Small is a 42-layer sparse Mixture-of-Experts model. According to the model card, each token is routed to six of 256 specialized experts, with two shared experts remaining active. That arrangement preserves a large pool of learned parameters while limiting how much of the network runs for each token.

The model accepts text, images and audio and produces text. Thinking Machines lists a context window of up to 1 million tokens at the model level, with 64,000-token and 256,000-token versions available through Tinker. Developers can download the full weights on Hugging Face under the Apache 2.0 license.

"Small" describes the comparison with Inkling. The standard BF16 checkpoint requires at least 600 GB of combined GPU memory, which Thinking Machines says can be supplied by four Nvidia B300 GPUs or eight H200 GPUs. A quantized NVFP4 checkpoint lowers the requirement to about 180 GB and can run on one B300 in W4A4 mode or two H200s in W4A16 mode. Those configurations put Inkling-Small in enterprise servers and cloud clusters, far outside the range of an ordinary laptop or developer workstation.

The reduction still changes the buying equation. A model that activates 12 billion parameters instead of 41 billion can lower inference costs and make self-hosting practical for a wider set of organizations with existing GPU capacity. It also gives Murati a model suited to coding assistants, retrieval systems, document analysis and specialized agents, where a company may value control and fine-tuning over the highest possible general benchmark score.

The benchmarks show a focused trade

VentureBeat reported that Artificial Analysis assigned Inkling-Small a score of 40 on its Intelligence Index, one point behind Inkling's 41. Thinking Machines' evaluation table reports stronger results for Inkling-Small on several coding and reasoning tests, including 80.2% on SWE-bench Verified against Inkling's 77.6%, and 64.7% on Terminal-Bench 2.1 against 63.8%.

Those figures require context. Thinking Machines says its SWE-bench result uses a bash-only harness and its Terminal-Bench score comes from an internal coding harness. The launch materials combine externally reported evaluations with internal testing, so the headline comparisons do not amount to independent validation across the full benchmark suite.

The smaller model also gives up ground in factual coverage. Inkling-Small scored 20.6% on SimpleQA Verified, less than half Inkling's 43.9%, and 15.5% on the tau3-Banking agent benchmark against 23.7% for Inkling. Those gaps make retrieval, source checking and human review important for factual or high-stakes uses.

Thinking Machines attributes the efficiency gains to changes in the pretraining data mix and machine-learning recipe. Inkling-Small began training after Inkling, giving Murati's researchers time to revise the process. Thinking Machines also used Inkling as a teacher for on-policy distillation of an earlier checkpoint, then continued agentic coding reinforcement learning for two weeks.

That sequence makes Inkling-Small a test of whether a frontier lab can turn the expense of one large model into a family of cheaper models. The flagship supplies training knowledge, while the smaller model broadens distribution and creates a more affordable entry point to Tinker.

Murati is tying open weights to paid customization

Thinking Machines is offering a limited-time 50% discount on the 64,000-token version through its Tinker pricing page. The listed rates are $0.58 per million prefill tokens, $1.44 per million sampled tokens and $1.73 per million training tokens. Cached prefill costs $0.116 per million tokens. The 256,000-token version carries higher rates.

The pricing reveals the commercial structure behind the release. Apache-licensed weights can spread Inkling-Small across external infrastructure and products. Tinker earns its place by handling training, scheduling and GPU management for developers who want a customized version without operating the cluster themselves. Thinking Machines can use an open-weight model to drive adoption while selling the managed work around adaptation.

Investors have already financed Murati to operate at frontier-lab scale. TechCrunch reported that Thinking Machines closed a $2 billion seed round led by Andreessen Horowitz at a $12 billion valuation in July 2025. In March 2026, Thinking Machines and Nvidia announced a multiyear partnership covering at least one gigawatt of Vera Rubin systems, alongside an undisclosed Nvidia investment.

Inkling-Small shows what Murati is buying with that capital and compute: a model family whose distribution can feed a training platform. Inkling demonstrated that Thinking Machines could train and release a large multimodal model. Inkling-Small makes the model considerably easier to operate while preserving enough performance to support real deployments. The next measure is whether developers adopt the weights, then choose Tinker when they need to make them their own.