Multimodal AI moved on several fronts this month, with Thinking Machines releasing a 975-billion-parameter open-weights model, researchers exposing gaps in how vision-language models handle infant-perspective video, and a new medical benchmark revealing that frontier models still fall short of online physicians on safety-sensitive clinical tasks.
Thinking Machines Releases Inkling, a 975B Open-Weights Multimodal Model
Thinking Machines — the AI startup founded by former OpenAI CTO Mira Murati — released Inkling on July 16 under an Apache 2.0 license, making its first major model freely available for commercial use. The 975-billion-parameter Mixture-of-Experts system processes text, images, and audio natively, and its weights are already downloadable on Hugging Face.
On third-party benchmarks, Inkling scores 77.6% on SWE-bench Verified — above NVIDIA Nemotron 3’s 71.9% — and 91.4% on VoiceBench, compared to 94.4% for Gemini 2.5 Pro on high reasoning effort, according to Thinking Machines’ announcement. Those figures place it below the current frontier but ahead of most open-weights rivals on both software engineering and voice understanding tasks.
Two design choices distinguish Inkling from competing open models. First, a “controllable thinking effort” mechanism lets operators dial compute up or down per query, trading cost against accuracy at inference time. Second, Thinking Machines states the model was built “to answer directly on topics that may be subject to censorship,
Sources
- Thinking Machines open sources first multimodal language model, Inkling, focused on low cost and ‘resistance to censorship’ – VentureBeat
- MedRealMM: A Real-World Multimodal Benchmark for Chinese Online Medical Consultation – arXiv AI
- NVIDIA and Japan Bring Full-Stack AI and Robotics to Every Industry – NVIDIA AI Blog
- AI Isn’t Smarter Than a Baby—Yet – Wired
- NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI – NVIDIA AI Blog
