Mira Murati (@miramurati)'s Thinking Machines Lab released Inkling-Small on July 30th, an open-weights multimodal model designed to deliver much of its larger sibling's reasoning and agentic performance with lower compute and inference costs.
In a six-post thread on X and a detailed technical announcement, Thinking Machines Lab said Inkling-Small has 276 billion total parameters and activates 12 billion for each token. The sparse Mixture-of-Experts model accepts text, images and audio, supports a context window of up to 1 million tokens and lets developers vary its reasoning effort to trade performance against cost.
Murati, OpenAI's former chief technology officer and brief interim CEO, launched Thinking Machines Lab publicly in February 2025 after leaving OpenAI to create what she described as "time and space" for her own exploration. Thinking Machines Lab subsequently raised about $2 billion at a $12 billion valuation in a seed round led by Andreessen Horowitz, with Nvidia, Accel, Cisco and AMD among the participating investors, WIRED reported.
Inkling-Small puts a more deployable model behind Murati's original pitch: AI systems that developers can understand and adapt instead of accessing solely through a closed provider's interface. The full model weights are available under the Apache 2.0 license, allowing commercial use, modification and local deployment subject to the license terms.
Smaller by design
Inkling-Small is roughly one-quarter the size of Inkling, which Thinking Machines Lab released on July 15th with 975 billion total parameters and 41 billion active. Both models use a sparse architecture that routes each token through a subset of the model's experts rather than activating every parameter.
Thinking Machines Lab says the smaller model benefited from lessons learned while training Inkling. Those changes included a revised pre-training data mix, a refined machine-learning recipe, on-policy distillation using Inkling as the teacher and another two weeks of reinforcement learning focused on agentic coding.
The resulting benchmark claims are strongest on coding and general reasoning. Thinking Machines Lab reports that Inkling-Small scored 80.2% on SWE-Bench Verified, 64.7% on Terminal-Bench 2.1 and 31.6% on the text-only version of Humanity's Last Exam. Inkling scored 77.6%, 63.8% and 29.7%, respectively, in the same table.
Those figures remain vendor-reported. Thinking Machines Lab used a bash-only harness for its SWE-Bench results and an internal coding harness for Terminal-Bench. The lab said it assigned zero scores to a small number of Terminal-Bench solutions contaminated by web search. It also acknowledged that Inkling retains an advantage in knowledge coverage and factuality. On SimpleQA Verified, Inkling-Small scored 20.6%, compared with Inkling's 43.9%.
The multimodal gap is narrower. Thinking Machines Lab reports scores of 74% on MMMU Pro and 90.1% on VoiceBench for Inkling-Small, close to Inkling's 73.5% and 91.4%. Images and audio are processed jointly with text rather than passed through a separate encoder, and the model can call Python to crop, enlarge and inspect portions of charts or documents while reasoning.
Tinker becomes the distribution channel
Thinking Machines Lab is releasing Inkling-Small through the same infrastructure it wants developers to use for customization. Murati introduced Tinker in October 2025 as a managed fine-tuning service that gives users control over training data and algorithms while Thinking Machines Lab handles distributed compute, scheduling and failure recovery.
Inkling-Small is available for fine-tuning through Tinker and for text, image and audio conversations through Tinker Playground. The model can also run locally through frameworks including Transformers, vLLM and SGLang, according to its model card.
Thinking Machines Lab's pricing page lists a limited-time 50% discount for the 64,000-token fine-tuning configuration. Rates are $0.58 per million prefill tokens, $1.44 per million sampled tokens and $1.73 per million training tokens. A 256,000-token configuration costs more, while the released model itself supports contexts of up to 1 million tokens.
That pairing makes Inkling-Small both a model release and a customer-acquisition vehicle for Tinker. Developers can inspect and host the weights themselves, then pay Thinking Machines Lab when they need managed fine-tuning or hosted inference. The approach places Thinking Machines Lab against open model publishers and customization providers at the same time.
The release also begins to show what Murati's unusually large seed round bought. Thinking Machines Lab has moved from a general thesis about customizable AI to a vertically connected product line: its own model family, a managed post-training service and a playground that can expose fine-tuned checkpoints directly to users. A gigawatt-scale Nvidia partnership announced in March gives Thinking Machines Lab a long-term route to the compute needed to keep expanding that stack.