Thinking Machines Lab Releases Inkling — A New Open-Weight AI Challenger
Thinking Machines Lab, the San Francisco-based AI startup founded by former OpenAI CTO Mira Murati, released its first in-house foundation model on July 15, 2026. Called Inkling, the model marks a major step in the company's mission to build AI that "extends human will and judgment."
Model Architecture
Inkling is a Mixture-of-Experts (MoE) transformer with: - 975 billion total parameters, 41 billion active - Up to 1 million tokens of context window - Pretrained on 45 trillion tokens of text, images, audio, and video - Controllable thinking effort to balance cost, latency, and performance
The architecture follows a DeepSeek-V3-style MoE design: 256 routed experts per layer with 2 shared experts, and 6 routed experts active per token.
A smaller variant, Inkling-Small, was also previewed with 276B total parameters (12B active) — achieving remarkably similar benchmark scores on reasoning and agentic tasks despite being roughly one-third the size.
Competitive Performance
Inkling delivers competitive performance against both open and closed models:
| Benchmark | Score | |-----------|-------| | AIME 2026 | 97.1% | | GPQA Diamond | 87.2% | | SWEBench Verified | 77.6% | | MMMU Pro | 73.5% | | VoiceBench | 91.4% | | Humanity's Last Exam | 29.7% |
On Design Arena's Agentic Web Dev leaderboard (blinded human evaluation), Inkling ranks among the strongest open-weights models with a score of 1257, close to GPT-5.6 Sol (1260) and Claude Opus 4.8 (1285).
Open-Weight Philosophy
Unlike proprietary models from OpenAI and Anthropic, Inkling is open-weight — users can download, run, customize, and fine-tune the full model. This positions it as a Western alternative to the popular open-source offerings from Chinese AI labs, which have seen surging adoption after Meta's shift to proprietary models.
Customization Platform: Tinker
Inkling is available for fine-tuning on Tinker, Thinking Machines' customization platform launched last October. The company is offering Inkling at a 50% discount for a limited time. The model can also be accessed via APIs on TogetherAI, Fireworks, Modal, Databricks, and Baseten.
Safety and Transparency
The company emphasized Inkling's safety credentials: it scores above 98% on StrongREJECT (a refusal test for harmful requests), and shows the strongest built-in safeguards of any open-weights model on the FORTRESS benchmark.
The Bigger Picture
Inkling arrives at a critical moment. The Western open-source AI ecosystem faces a gap after Meta's retreat from open-weight releases, and businesses have increasingly turned to Chinese models. Inkling — with its combination of openness, multimodality, controllable efficiency, and safety — aims to fill that void.
