AI Research Roundup: July 2026’s Biggest Moves

Three developments dominated AI research and applied science in July 2026: Thinking Machines released its 975-billion-parameter open-weights model Inkling, OpenAI researcher Miles Wang is leaving to launch a drug discovery startup targeting a $2 billion valuation, and PsiQuantum detailed its photonic quantum computing architecture in a new technical disclosure. Together, they illustrate how AI research is rapidly moving from benchmarks to physical and biological applications.

Thinking Machines Releases Inkling, a 975B Open-Weights Model

Thinking Machines — founded by former OpenAI CTO Mira Murati — released Inkling on July 16, 2026, under an Apache 2.0 license, making it freely available for commercial and enterprise use. The model scores 77.6% on SWE-bench Verified, outperforming NVIDIA’s Nemotron 3 (71.9%), and 91.4% on VoiceBench, according to Thinking Machines’ release announcement.

Inkling is a natively multimodal Mixture-of-Experts (MoE) system capable of processing text, images, and audio. Its 975 billion total parameters are paired with a “controllable thinking effort” mechanism that lets operators trade compute cost against output quality at inference time — a design choice aimed at enterprise workloads where predictable cost matters as much as raw performance.

A smaller companion model, Inkling-Small at 276 billion parameters, was announced in preview for latency-sensitive deployments. Both models are available on Hugging Face and through Thinking Machines’ own Tinker API.

One differentiator the company highlighted explicitly: Inkling was designed, in the company’s words, “to answer directly on topics that may be subject to censorship” — positioning it for enterprises that want factual outputs regardless of content sensitivity. Horace He, a researcher at Thinking Machines previously from PyTorch, noted on X that the engineering execution behind the release was a significant collective effort.

OpenAI Researcher Leaves to Build AI Drug Discovery Startup

Miles Wang, an OpenAI researcher whose published work includes using AI to accelerate biological discovery, is leaving the company to found a startup focused on AI models for drug discovery, according to TechCrunch. Wang is in talks to raise approximately $200 million at a $2 billion valuation, with Lightspeed in discussions to lead the round, per four people with knowledge of his plans. Wang disputed the funding figures reported by TechCrunch but did not provide corrected numbers; Lightspeed did not respond to comment requests.

Wang joined OpenAI in 2024 after leaving Harvard, where he was pursuing a computer science degree. At OpenAI, he co-authored research papers evaluating how AI models can automate and accelerate scientific workflows.

Two sources told TechCrunch the new startup may focus on drug repurposing — finding new therapeutic uses for FDA-approved compounds or drugs that previously failed trials. That approach can compress time-to-revenue significantly compared to de novo drug development, since safety data already exists.

The Broader Drug Discovery Investment Wave

Wang’s departure tracks a surge in venture capital flowing into AI-driven life sciences. Chai Discovery, a two-year-old startup that predicts molecular interactions to identify new drugs, raised $400 million at a $3.8 billion valuation this week. Its co-founder Josh Meier also passed through OpenAI as a researcher. Separately, Google DeepMind spinout Isomorphic Labs closed a $2.1 billion Series B in May 2026, also targeting AI-native drug discovery, according to TechCrunch.

Singularity Gate Benchmark Tests AI’s Predictive Scientific Reach

A benchmark called the Singularity Gate — designed to test whether frontier AI models can predict paradigm-breaking scientific discoveries published after their training cutoff — has published new results for several leading models, according to a Reddit Singularity post summarizing the evaluation.

Claude Fable 5 currently leads the benchmark. However, the original Fable 5 responded to 45% of benchmark tasks, while the latest version dropped to 39%, with the post noting both a higher refusal rate and a slight degradation in performance on tasks both versions attempted. GPT-5.6 Sol showed a measurable improvement over GPT-5.5 on the same evaluation.

The benchmark’s methodology — testing predictions against discoveries made after a model’s knowledge cutoff — is designed to measure genuine extrapolative reasoning rather than memorized answers. Results should be interpreted with caution: the Reddit post is a community summary, not a peer-reviewed publication, and the benchmark’s scoring methodology has not been independently audited.

PsiQuantum’s Photonic Quantum Architecture Targets Drug Simulation

PsiQuantum, founded in 2016 by four physicists from UK universities, has detailed its plan to build a fault-tolerant quantum computer using photons rather than superconducting qubits, according to MIT Technology Review. The planned system would occupy roughly 100 stainless-steel cabinets, each cooled to near absolute zero with liquid helium, housing chips through which individual photons travel through optical switches and beam splitters.

The machine does not yet exist in operational form, but PsiQuantum has identified a concrete target application: simulating cytochrome P450 enzymes, which metabolize drugs in the human body. Philipp Ernst, a vice president at PsiQuantum, told MIT Technology Review that estimating enzyme behavior for a specific drug currently takes over 10 years with conventional methods — a timeline the company believes quantum simulation could compress substantially.

PsiQuantum’s photonic approach differs from superconducting competitors like IBM and Google: photons are less prone to certain decoherence mechanisms, but manufacturing the required optical chips at scale remains an unsolved engineering challenge.

What This Means

The convergence of AI and life sciences research is now attracting capital at a scale that rivals core AI infrastructure investment. Three separate drug discovery ventures — Wang’s unnamed startup, Chai Discovery, and Isomorphic Labs — collectively represent over $6 billion in recent funding, all within months of each other. That capital concentration suggests investors believe the next commercially valuable AI application is molecular biology, not just software.

The Inkling release from Thinking Machines adds a credible open-weights option to enterprise AI deployments, particularly for organizations that cannot use closed APIs for compliance or data-sensitivity reasons. Its explicit anti-censorship design is a deliberate market signal, differentiating it from models tuned toward conservative refusal behavior.

The Singularity Gate benchmark, while not peer-reviewed, points to a genuine research question: can large language models extrapolate beyond their training data to anticipate scientific results? The performance drop between Fable 5 versions — from 45% to 39% response rate — suggests that post-training alignment changes can degrade research-relevant capabilities even when general benchmarks improve, a tradeoff the field has not resolved.

PsiQuantum’s photonic roadmap remains the longest bet in this group. Quantum advantage for drug simulation is theoretically compelling but practically years away, and the company has yet to demonstrate a fault-tolerant system at any scale.

FAQ

What is the Singularity Gate benchmark?

The Singularity Gate is an evaluation designed to test whether frontier AI models can predict major scientific discoveries published after their training data cutoff. It measures extrapolative reasoning rather than memorized knowledge, though its methodology has not been independently peer-reviewed.

What is drug repurposing in the context of AI research?

Drug repurposing means identifying new therapeutic uses for compounds already approved by the FDA or that have existing safety data from prior trials. AI models can accelerate this by scanning molecular interaction data at scale, and the approach reduces development timelines because safety testing is already complete.

How does PsiQuantum’s photonic quantum computer differ from superconducting designs?

PsiQuantum uses photons — particles of light — as qubits rather than the superconducting circuits used by IBM and Google. Photons are naturally less susceptible to certain types of quantum decoherence, but manufacturing the precision optical chips required at production scale remains a major unsolved engineering challenge.