Mira Murati’s Lab Enters the Open-Weights Race
Thinking Machines Lab, the startup founded in early 2025 by former OpenAI chief technology officer Mira Murati, released its first public model on July 15, 2026. Inkling is a 975-billion-parameter multimodal system, licensed under Apache 2.0, and it arrives with a framing that stands out against a year of leaderboard-chasing announcements: Thinking Machines says openly that Inkling is not the strongest model available today, open or closed, and is not trying to be.
That candor is itself the story. Rather than compete on raw capability against Claude Fable 5, GPT-5.6 Sol, or Gemini 3.1 Pro, Inkling is positioned as the best available open foundation for organizations that want to build their own specialized model rather than rent access to someone else’s. It is also, notably, the largest open-weight model to come out of the United States so far, entering a category that had been dominated almost entirely by Chinese labs — DeepSeek, Moonshot, and Zhipu among them.
- Release date: July 15, 2026
- License: Apache 2.0 — full commercial use, modification, and redistribution permitted
- Architecture: Sparse Mixture-of-Experts transformer, 975 billion total parameters, approximately 41 billion active per token (six experts activate per token)
- Training data: 45 trillion tokens spanning text, images, audio, and video
- Inputs: Text, images, and audio (16kHz WAV, best under roughly 20 minutes); output is text only
- Context window: Up to 1 million tokens
- Companion release: Inkling-Small, a 276-billion-parameter (12 billion active) sibling still being finalized, with weights promised once testing is complete
Why the Audio Input Is the Detail Worth Noticing
Buried underneath the parameter count is a feature that several independent reviewers flagged as the release’s most distinctive capability: native audio understanding. Inkling accepts spoken audio directly as an input alongside text and images, which one widely circulated launch-day comment described as making it the largest open-weight model that supports audio input. That matters for a specific and immediately useful class of applications — a model that can listen to a recorded support call, a meeting, or a voice memo directly, rather than requiring a separate transcription step before any reasoning can happen.
Combined with the million-token context window, that gives Inkling a genuine practical edge for workflows involving long recorded audio, even though the model does not lead the field on raw intelligence benchmarks.
The Benchmark Picture: Strong, Not Dominant
Thinking Machines published a detailed set of benchmark comparisons at launch, and the honest read is a mixed one. Inkling scored 97.1% on the 2026 American Invitational Mathematics Examination benchmark and 87.2% on GPQA Diamond, a graduate-level scientific reasoning test. On SWE-bench Verified it reached 77.6%, and on the more demanding SWE-bench Pro Public it scored 54.3%. On Humanity’s Last Exam, Inkling posted 29.7% without tool access and 46.0% with tools enabled — competitive against other open models, but below the strongest closed frontier systems in Thinking Machines’ own comparison table.
Measured against its closest open-weight rivals, the picture is one of frequent parity rather than outright superiority. Against DeepSeek V4 Pro, Inkling trails on SWE-bench Verified (77.6% versus 80.6%) and on the factuality benchmark SimpleQA Verified (43.9% versus 57.0%), but overtakes it on mathematical reasoning, scoring 97.1% on AIME 2026 against DeepSeek’s 96.7%. Against Kimi K2.6, Moonshot’s prior-generation open model, Inkling lags on GPQA Diamond (87.9% versus 91.1%), on the web-browsing benchmark BrowseComp (77.1% versus 83.2%), and on Humanity’s Last Exam with tools (46.0% versus 54.0%), but leads on general instruction-following, scoring 79.8% on IFBench against K2.6’s 76.0%.
Inkling is not the smartest model you can get today. What makes it interesting is that it is a rare, credible American open-weights model that natively understands audio — and one you can actually download and run yourself.
On Artificial Analysis’s independent Intelligence Index, Inkling scores 41, placing it 10th of 97 models tracked — a respectable mid-pack result for a first release, well short of the frontier tier occupied by Fable 5, Opus 5, and GPT-5.6 Sol, but ahead of a meaningful share of both open and closed competitors.
The Pricing Paradox
One detail that surprised several early reviewers: despite being released under a fully permissive open-source license, Inkling is priced on Thinking Machines’ own hosted API like a frontier-tier model, not like a budget open-weight option. That has led to an odd bifurcation in how the model gets used in practice. Teams that want to self-host — running Inkling on their own infrastructure via the weights posted on Hugging Face, including an NVFP4 checkpoint tuned for current-generation NVIDIA hardware — get a genuinely competitive open model at effectively zero marginal licensing cost. Teams that would rather rent access through Thinking Machines’ own Tinker platform pay closer to what a closed frontier model would charge, which undercuts the traditional value proposition of choosing an open-weight release in the first place.
The practical recommendation emerging from early developer coverage is straightforward: Inkling is a strong pick if you intend to own the weights, need native audio or extended context, and have the infrastructure to self-host. It is a weaker pick if your goal is simply the cheapest available API access or the single highest benchmark score on the market — for either of those goals, other releases from July 2026 are more directly suited.
What the Documentation Does and Doesn’t Say
Thinking Machines published a model card alongside a separate Training Data Documentation page, both notably shorter than the documentation that has become standard from larger US labs. The training data document states that the datasets used to develop Thinking Machines’ models include content that is in the public domain as well as content that may be subject to intellectual property protection — a fairly minimal disclosure that leaves most specifics about sourcing unaddressed. That brevity has drawn some comment from developers accustomed to more detailed data-provenance disclosures from larger labs, though it has not meaningfully slowed adoption.
Positioning Inkling Within the Broader Open-Weights Landscape
Inkling’s arrival gives the US open-weights ecosystem a genuinely new frontier-scale contender alongside NVIDIA’s Nemotron line and Google’s Gemma 4, entering a category where Chinese labs — DeepSeek, Moonshot, and Zhipu among them — have set much of the pace over the past year. Whether Inkling meaningfully shifts that balance will depend less on its launch-week benchmark scores than on what developers actually build once they start fine-tuning it through Tinker, Thinking Machines’ dedicated training platform.
Tinker: The Platform the Model Is Built Around
Inkling does not exist in isolation from the rest of Thinking Machines’ product strategy. The company positions its Tinker platform — a dedicated fine-tuning and training environment — as the primary way most organizations are expected to interact with the model in practice, rather than treating Inkling as a finished, standalone chatbot. That framing helps explain some of the choices that look unusual next to other 2026 releases: the comparatively short model card, the minimal training-data documentation, and the pricing structure on Thinking Machines’ own hosted API that reads as frontier-tier rather than budget-open-weight.
None of those choices make much sense if Inkling is judged purely as a finished consumer product competing head-to-head with Claude Opus 5 or GPT-5.6 Sol on a general leaderboard. They make considerably more sense once Inkling is understood as raw material — a capable, permissively licensed, natively multimodal starting point that Tinker is designed to help developers reshape into something narrower and more specialized, whether that is a customer-support agent tuned on a company’s own audio call logs, a coding assistant fine-tuned on a specific internal codebase, or a domain-specific research tool built around a particular scientific literature corpus.
The “Resistance to Censorship” Framing, and What It Actually Means
Some launch-week coverage described Inkling’s positioning in terms of resistance to censorship — a framing tied less to any single benchmark and more to the practical consequences of Apache 2.0 licensing itself. Because the license places no restriction on how a downstream deployer configures the model’s behavior once it is running on their own infrastructure, organizations that fine-tune Inkling retain full control over content policies, safety filtering, and refusal behavior in their own deployment, rather than inheriting a fixed set of guardrails baked in by the original lab and enforced through a hosted API. That is a structural property of any Apache 2.0-licensed open-weight model, not something unique to Inkling’s training process — but Thinking Machines’ choice to release at this parameter scale under this particular license puts a meaningfully more capable option into that category of the market than had previously been available from a US lab.
That is, in fact, precisely the bet the lab is making explicit: Inkling is not pitched as the model to win today’s leaderboard, but as the open base other people’s specialized models get built on top of tomorrow. Whether that bet pays off is likely to become clearer only in the months after launch, as fine-tuned derivatives built on Inkling’s weights start appearing in production — a test that, unlike a benchmark score, cannot be settled on release day.