Mira Murati's Lab Previews Inkling-Small, a Lean Open Model Built for Builders on a Budget
News Summary
Thinking Machines Lab, the AI research company founded by former OpenAI chief technology officer Mira Murati, unveiled its first open-weight model family on July 15, 2026, at 9:00 AM Pacific Time. The flagship release, called Inkling, is a 975-billion-parameter multimodal mixture-of-experts model, and it arrives alongside a lighter sibling, Inkling-Small, aimed at teams that want strong reasoning performance without access to large GPU clusters. Rather than chasing the top spot on leaderboards, the company positioned both models as flexible starting points for developers who want to customize and fine-tune AI systems for their own products.
What Inkling and Inkling-Small Are
Inkling is built on a mixture-of-experts transformer architecture with 256 routed experts and two shared experts per layer, of which six routed experts activate for any given token. That design gives the full model 975 billion total parameters while only about 41 billion are active during inference, keeping compute costs manageable. Inkling-Small follows the same architecture at a reduced scale: 276 billion total parameters with roughly 12 billion active per token. Both models support context windows of up to 1 million tokens and were trained on a combined 45 trillion tokens spanning text, images, audio, and video.
Notably, Inkling-Small was shared as a technical preview at launch. Its full downloadable weights are being held back briefly for additional safety and quality testing, though developers can already fine-tune it through Thinking Machines' customization platform, Tinker.
A "Thinking Effort" Dial for Cost Control
One of the more distinctive features across the Inkling family is a controllable "thinking effort" setting, adjustable through the system message, which lets developers trade reasoning depth for speed and token cost. Company researchers reported that this approach let Inkling match rival systems on agentic coding benchmarks such as Terminal Bench 2.1 while using roughly one-third the tokens of Nvidia's Nemotron 3 Ultra model. For Inkling-Small specifically, the efficient-reasoning design is aimed squarely at latency-sensitive and budget-conscious workloads, including coding assistants, automated grading systems, and generating synthetic training data for smaller downstream models.
Benchmark Performance
At its highest reasoning-effort setting, Inkling posted an 87.2 percent score on GPQA Diamond, 97.1 percent on AIME 2026, and 77.6 percent on SWE-Bench Verified, a widely used measure of real-world coding capability. On multimodal evaluations, the model reached 73.5 percent on MMMU Pro for vision tasks and 91.4 percent on VoiceBench for audio understanding. Thinking Machines reported that Inkling-Small trails only modestly behind its larger counterpart on many of these same evaluations, with vision and audio scores typically within one to three percentage points of the full-size model, despite using less than a third of the active parameters.
Training Approach and Safety Work
The Inkling family was trained using a hybrid optimization method, applying the Muon optimizer to large weight matrices and Adam to the remaining parameters, followed by supervised fine-tuning and large-scale reinforcement learning involving more than 30 million rollouts. The company said it placed particular emphasis on calibration, training the models against real-world questions with resolved outcomes so that they express appropriate uncertainty rather than confidently guessing. Safety evaluations included adversarial red-teaming benchmarks, where Inkling scored 98.6 percent on the StrongREJECT test and 78.0 percent on the FORTRESS adversarial safety suite, alongside external review from independent testers.
Availability and Pricing
Inkling's full weights are downloadable now through Hugging Face, including a compressed NVFP4 checkpoint optimized for Nvidia's Blackwell chips, and the model is also accessible through hosting partners including Together AI, Fireworks, Modal, Databricks, and Baseten. Thinking Machines is offering a limited-time 50 percent discount for developers fine-tuning through Tinker, along with free, time-limited access to an Inkling playground inside the Tinker console. Inkling-Small's full weights are expected to follow once additional testing concludes.
Industry Context
The release adds a second entrant to Thinking Machines' public product lineup, following the earlier debut of the Tinker fine-tuning platform. Company researchers were candid that Inkling is not intended to be the single strongest model available, open or closed; instead, the emphasis is on giving developers, universities, and research labs a well-rounded, efficient foundation they can adapt for specialized science, coding, and reasoning applications. Early enterprise interest was highlighted through a collaboration with investment firm Bridgewater Associates, which reported strong results applying Inkling to financial-reasoning benchmarks at a fraction of the computing cost of comparable proprietary models.