Mira Murati’s Inkling Challenges Closed AI Giants With Open Multimodal Power

Former OpenAI chief technology officer Mira Murati left the company in early 2025 to launch Thinking Machines Lab. She brought along a cluster of top researchers. The new venture quickly raised $2 billion in seed funding and reached a $12 billion valuation within months. Now the startup has released Inkling. This open-weights model arrives as a direct bid to shift power away from proprietary systems.

Inkling packs 975 billion total parameters yet activates just 41 billion at any moment. The mixture-of-experts design keeps compute demands manageable. It handles a full 1 million token context window. And it processes text, images and audio from the start. The model was pretrained on 45 trillion tokens spanning those same modalities plus video. Details come straight from the company’s announcement on its site.

But raw size tells only part of the story. Inkling stands out for controllable effort. Users can dial reasoning intensity up or down per task. Spend fewer tokens on routine steps. Pour more compute into tough problems. The approach yields strong results on agentic benchmarks while trimming costs. On Terminal Bench it matches heavier models at one-third the token count. Thinking Machines shared those figures in its July 15, 2026, post.

Performance across domains looks competitive. The model scores well on math, science and knowledge tests. It shines in coding environments where it writes programs, calls tools and iterates. Instruction following stays reliable even across long, complex prompts. Factuality improves through targeted reinforcement learning. And calibration feels tight. The system reports confidence levels that match its actual accuracy on forecasting tasks.

Multimodal handling pushes further. Inkling transcribes speech, reasons over audio clips and describes charts or photos. It avoids separate encoders. Images arrive as patches. Audio becomes spectrograms. Early tests on VoiceBench, MMAU and AudioMC place it near the front of open-source packs. Vision tasks benefit from built-in Python tools for cropping or zooming. The company highlighted these strengths in the same announcement that SiliconANGLE covered in related interaction model coverage.

Safety receives serious attention. Inkling refuses harmful requests without over-refusing benign ones. It scores above 98 percent on StrongREJECT. On the FORTRESS benchmark it leads every open-weights peer. Training included rubric-based graders that penalize hallucinations. The result is a system less prone to fabricating details. Yet it resists heavy censorship. “We trained Inkling to answer directly on topics that may be subject to censorship,” the team stated in its release notes.

That stance fits the broader philosophy. Murati and her co-founders, including John Schulman, Barrett Zoph, Lilian Weng and others from OpenAI, set out to build systems that extend human judgment rather than replace it. They structured Thinking Machines as a public benefit corporation. Murati holds voting control. The goal centers on easy customization so organizations can adapt models to their own data, values and workflows. A Built In profile from April 2026 captured the mission in detail.

Tinker, the company’s fine-tuning platform, makes that adaptation practical. Launched in late 2025, it lets developers specialize models without massive distributed training clusters. Inkling integrates immediately. Users get a 50 percent discount for a limited window. Context windows of 64,000 or 256,000 tokens are available now. A new playground inside Tinker lets teams chat with the model, run agentic searches and test ideas before committing resources. The announcement lists integrations with Together AI, Fireworks, Modal, Databricks, vLLM, llama.cpp and Hugging Face.

Community reaction on X lit up within hours of the July 15 debut. One post noted the model does not top every benchmark yet carves a niche around human-AI collaboration. Another highlighted its resistance to censorship and customization potential. Chinese open models had filled much of the Western vacuum over the past year. Inkling marks a serious American entry. Enterprises gain control over data and costs instead of depending solely on API providers. A thread from @lu__jasper captured that sentiment clearly.

Comparisons matter. Inkling does not dominate every leaderboard. Closed models from OpenAI, Anthropic and Google still hold edges in some areas. Yet the open release changes the equation. Developers can inspect weights, run the system on their hardware and build derivatives. Hugging Face hosts both the original and an optimized NVFP4 version for NVIDIA Blackwell chips. That accessibility echoes earlier moves by Meta with Llama but adds native multimodality and controllable reasoning from day one.

The release builds on prior work. Thinking Machines first offered Tinker for adapting existing open models. It previewed interaction models in May 2026 that cut response latency to under 0.4 seconds for more natural conversation. Those systems use a dual-model setup. A small fast component handles dialogue while a larger one tackles heavy reasoning in the background. SiliconANGLE reported the preview and quoted analysts who see demand for AI that matches human turn-taking rhythms. Inkling shares architectural DNA with that effort. Its Inkling-Small variant, at 276 billion total parameters and 12 billion active, already shows promise for lower-latency use cases.

Training followed an unusual path. The team mixed Muon and Adam optimizers. They drew hyperparameter insights from research on modular manifolds. Post-training mixed supervised fine-tuning bootstrapped from other strong models with large-scale reinforcement learning. Over 30 million rollouts produced steady gains in reasoning quality. Chain-of-thought patterns began to compress naturally. Effort control emerged through system prompts that set token budgets. The company detailed the process in its technical notes.

Epistemics formed a core focus. The model learns to abstain when uncertain. It forecasts future events with calibrated confidence. Reinforcement learning used claim graders to reduce hallucinations. These steps produce a system that feels more trustworthy for high-stakes work. Financial tasks, scientific analysis and legal review could benefit. A June 2026 company note on replicating expert judgment in finance hinted at the direction.

Critics point to the funding frenzy. Some observers called the $2 billion round remarkable for a company that had not yet shipped a flagship model. A YouTube analysis from Wall Street Millennial questioned the valuation relative to delivered products. Yet talent density remains hard to ignore. The founding team carries direct experience from GPT-4 development and beyond. That pedigree helped attract further researchers even after reported poaching attempts by OpenAI itself.

Practical demonstrations impress. Inkling built a one-shot web application complete with embedded browser. It climbed leaderboards on Design Arena. It refined a multiplayer game through iterative loops. These examples go beyond static benchmarks. They show an agent that can plan, execute and correct course. Controllable effort lets developers trade speed for quality on demand. The combination could prove more useful in production than models that always run at maximum intensity.

Future releases will expand the family. Thinking Machines calls Inkling the first in a series. Later versions may close more of the gap with closed frontier systems. The company continues to publish research on distillation, inference efficiency and interaction paradigms. Its blog hosts papers on on-policy distillation and LoRA improvements. That openness aligns with the public benefit charter.

Enterprises face a choice. Stick with polished API services that improve weekly but lock data inside someone else’s infrastructure. Or download weights, fine-tune locally or in controlled clouds and retain full governance. Inkling does not win every head-to-head test. Its balanced profile, multimodal reach and customization pathway still make it compelling. Early adopters on X already discuss integration plans for coding agents, content pipelines and domain-specific assistants.

The timing feels deliberate. Western open model momentum had slowed. Chinese labs filled the gap with strong, uncensored alternatives. Inkling reasserts an American presence that prioritizes transparency and control. Whether it sparks a wave of derivative models depends on how quickly the community experiments. The weights are public. The playground is live. Tinker stands ready. The experiment has begun.