Mira Murati's Thinking Machines Unleashes Inkling, a 975B-Parameter Open-Weights Challenger to Closed AI
The most consequential AI story of the day is not another closed-model upgrade from a frontier lab. It is the arrival of **Inkling**, a 975-billion-parameter open-weights model released by **Thinking Machines Lab**, the startup founded by former OpenAI CTO Mira Murati in early 20
The most consequential AI story of the day is not another closed-model upgrade from a frontier lab. It is the arrival of Inkling, a 975-billion-parameter open-weights model released by Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati in early 2025 [1][2]. After raising a record $2 billion seed round at a $12 billion valuation before shipping any product, the company has finally put its first model into the wild—and it is making a bet that directly contradicts the strategy Murati once helped execute at OpenAI [4].
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active parameters per token, built around 256 routed experts and two shared experts [1][3]. It was trained from scratch on 45 trillion tokens spanning text, images, audio, and video using Nvidia GB300 NVL72 systems, and it supports a context window of up to one million tokens [1][2]. The model is natively multimodal, accepts text, image, and audio inputs, and is released under the permissive Apache 2.0 license, with weights downloadable from Hugging Face [2][3]. That makes it the largest American open-weights model to date and places it in the same tier as Chinese open-weights leaders such as DeepSeek V4, GLM 5.2, and Kimi K2.6 [1].
The hardware requirements are sobering. Running the native BF16 checkpoint demands more than two terabytes of aggregated GPU memory—roughly eight Nvidia B300 accelerators or sixteen H200s [1][3]. A quantized NVFP4 checkpoint lowers the bar to about 600 GB, runnable on four B300s or eight H200s [3]. For most developers, the practical path will be API access through Thinking Machines' Tinker fine-tuning platform or third-party inference providers such as TogetherAI, Fireworks, Modal, Databricks, and Baseten [1][4].
What matters most here is not raw benchmark supremacy. Thinking Machines is explicit that Inkling is "not the strongest overall model available today, open or closed" [2]. Instead, the company is selling customizability. Inkling is designed to be fine-tuned on proprietary data, deployed on infrastructure the customer controls, and tuned for specific workflows rather than rented by the token through a closed API [2][4]. That positioning lands squarely in the middle of an intensifying enterprise debate: as the performance gap between top models narrows, do companies want the smartest general-purpose model, or the one they can make their own?
The open-versus-closed divide is becoming a defining commercial fault line in AI. Proprietary models from OpenAI, Anthropic, and Google typically run behind APIs and charge per token, which can scale unpredictably for high-volume enterprise workloads [5]. Open-weight models shift the cost structure toward owned infrastructure and give organizations control over data governance, fine-tuning, and integration—advantages that matter heavily in finance, healthcare, and government [5]. Palantir CEO Alex Karp recently argued on CNBC that frontier closed-model tools are too expensive and lack clarity on intellectual-property protections, a sentiment that appears to be pushing more enterprises toward open-weight alternatives [4][5]. Stanford University's AI Index Report 2026 notes that "leading models are now nearly indistinguishable from one another" and that "open-weight models are more competitive than ever" [5].
Murati's move carries particular symbolism. She was at OpenAI in 2019 when the lab, originally founded on a promise of openness, withheld the full version of GPT-2 over misuse fears—a decision that marked the beginning of its retreat from fully open releases [4]. Now she is releasing a frontier-class model with open weights under Apache 2.0, a license that lets end users freely fine-tune and deploy the model for their own purposes [1][2]. The Register's framing is pointed: this is "what Altman won't" do [1]. Whether that openness persists across future Thinking Machines models is an open question; Axios notes that Murati may apply the same case-by-case logic she learned at OpenAI, releasing openly when risks feel manageable and holding back when they do not [4].
There are caveats. Benchmark claims should be treated skeptically—"gaming AI benchmarks isn't exactly difficult," as The Register puts it—and Inkling's own charts show it trailing proprietary models such as Anthropic's Claude and OpenAI's GPT [1]. The model also used synthetic data generated by existing open models, including Moonshot AI's Kimi K2.5, in its final training phase, which raises familiar questions about the provenance and quality of training data in the open-weights ecosystem [4]. And the hardware requirements mean that, despite the open license, only well-capitalized organizations or cloud providers can realistically run the full model on-premise.
Still, Inkling is more than a product launch. It is a strategic argument made manifest: that the next phase of AI competition will be fought not only over which model scores highest on a leaderboard, but over which model gives enterprises the freedom to adapt, own, and control their AI stack. With the U.S.-China model-performance gap effectively closed and open-weights alternatives gaining ground, Thinking Machines is betting that customization and cost predictability will matter more than marginal benchmark gains [5].
If that bet is right, Inkling will be remembered as the moment a credible American alternative to both closed frontier labs and Chinese open-weights leaders finally entered the arena. If it is wrong, it will still have forced the industry to confront a question it can no longer avoid: in a world where the best models are increasingly similar, why should enterprises keep renting what they could own?
Synthesizer fusing final answer…
title: "Strait of Hormuz on the Brink: US-Iran Strikes Reignite a War That Never Really Ended" date: 2026-07-16 category: "world" tags: ["Iran", "United States", "Strait of Hormuz", "Oil", "Middle East", "Geopolitics"] sources: ["https://apnews.com/article/iran-us-hormuz-strait-war-july-16-2026-f98ff56554de2336f0e85bb5fdcae769", "https://www.bbc.co.uk/news/articles/c2lq1ed28jxo", "https://www.bbc.com/news/articles/cy0608wy8pro", "https://www.channelnewsasia.com/business/oil-prices-rise-4th-day-us-strikes-iran-raise-fears-wider-conflict-6257256"]
Strait of Hormuz on the Brink: US-Iran Strikes Reignite a War That Never Really Ended
The war that Washington and Tehran officially paused in June is collapsing back into open combat. In the past 72 hours, the United States has expanded strikes into northern Iran for the first time in the current wave of hostilities, hit targets around Tehran itself, disabled an oil tanker trying to break a freshly reimposed naval blockade, and absorbed Iranian missile and drone fire directed at bases in Bahrain, Jordan, and Kuwait [1][2]. Brent crude, after four consecutive sessions of gains, traded above $85 a barrel on Thursday — more than 15% above its pre-war level — as tanker traffic through the Strait of Hormuz fell to a two-month low [1][4].
What is unfolding is not a new conflict. It is the slow, predictable detonation of an interim deal that was always too thin to hold.
The truce that was already dead
The June memorandum of understanding was supposed to end hostilities that began when the US and Israel launched the war on February 28. In its early days, Iran effectively closed the Strait of Hormuz, sending oil, fertilizer, and shipping rates soaring and giving Tehran extraordinary leverage [1]. The June deal traded that leverage for an easing of the US naval blockade — but it left the most contentious issue, control of the strait itself, unresolved.
By Monday, that ambiguity had caught fire. President Donald Trump declared the US the "guardian" of the Strait of Hormuz and announced a 20% reimbursement fee on all cargo passing through it, a move Iran described as a unilateral dismantlement of the truce [3]. Within 24 hours, Trump reversed the fee under pressure from Gulf leaders, replacing it with vague promises of "massive" trade and investment deals — and simultaneously reimposed the naval blockade on Iranian ports [3]. Tehran's response was to attack seven commercial ships in the strait, killing or wounding nearly a dozen crew members according to US Central Command, and to strike two UAE-flagged tankers, killing one Indian sailor [3].
A wider battlefield, deliberately chosen
The new strikes are notable for what they target. For the first time in this round, US aircraft hit areas around Tehran and struck Semnan province, home to Iran's ballistic missile production and space launch infrastructure [1][2]. An attack on Greater Tunb Island, a strategic outpost in the strait, went after Iranian defense and missile sites [1]. Another strike hit a barracks of Iran's 388th Mechanized Infantry Brigade in Sistan and Baluchestan province, reportedly killing seven soldiers [1].
This is escalation by geography. Earlier rounds focused on coastal surveillance and IRGC fast-boat facilities. By reaching into Iran's missile heartland and the capital's outskirts, the US is signaling that the blockade is no longer a pressure tool — it is the opening phase of something larger. Trump's own framing reinforces the read: "Next week it gets really bad for them. We're going to knock out all their power plants. We're going to knock out all their bridges unless they get to the table and negotiate" [3].
Iran's parliament speaker, Mohammad Bagher Qalibaf, responded by saying Tehran was prepared for a "fuller military confrontation," while the Islamic Revolutionary Guard Corps warned that "the export of oil and gas from the region will be either for everyone or for no one" [1]. That statement is a threat aimed less at the US Navy than at Gulf energy infrastructure — Saudi, Emirati, and Iraqi pipelines and terminals that sit in Iran's missile range. It is also a reminder that Iran retains options its enemies would prefer to forget, including a long-promised closure of the Bab el-Mandeb strait via its Houthi partners, which would put a second vital energy artery at risk [4].
Why this round is different
Three things make the current flare-up more dangerous than the previous rounds.
First, the negotiating track has effectively collapsed. Trump's "better behave or else" rhetoric, combined with Qalibaf's claim that Iran has "no reason" to abide by a deal that doesn't benefit it, leaves no off-ramp that both sides can claim as a win [2][3]. The earlier rounds were punctuated by mediators — Oman, Qatar, Iraq — who could shuttle messages. This round has Trump publicly musing that the alternative to a deal is to "just finish it off" [1].
Second, the civilian-infrastructure threat has been put on the table explicitly. Trump has openly threatened Iran's bridges and power plants, a category of target the UN human rights chief has already warned could constitute a war crime under the Geneva Conventions [3]. Once that line is named in a presidential primetime interview, the legal and moral floor of the conflict drops.
Third, the economic pressure is now bidirectional in a way it wasn't before. Brent crude above $85 is hurting the global economy and the Republican Party's November prospects, but it is also providing Iran with revenue from every barrel that still moves. Goldman Sachs has warned Brent could exceed $110 a barrel in the fourth quarter if Gulf export recovery continues to stall, while ING notes that US commercial oil inventories are at their lowest seasonal level since 2018, leaving the market unusually vulnerable to any further supply shock [4].
What to watch next
The next 72 hours will likely determine whether this is another punishing cycle of escalation-and-de-escalation, or the prelude to a full regional war. Three indicators matter most.
Shipping through the strait. On Wednesday, just seven vessels transited, down from thirteen the day before [4]. If that number collapses toward zero, insurers will withdraw coverage, and the war will enter a phase where oil markets price in a multi-month disruption rather than a temporary premium.
Gulf state responses. Kuwait and Bahrain intercepted Iranian drones overnight; the UAE lost a crew member to an Iranian cruise missile strike earlier in the week [1][3]. If Gulf states begin treating Iranian retaliation as a direct attack on their sovereignty rather than collateral damage, the diplomatic isolation of Tehran that has underwritten US strategy could fracture.
The Houthi card. Analysts cited in Asian markets are already pricing in the possibility that Iran will activate its proxies in Yemen to threaten the Bab el-Mandeb [4]. A second chokepoint closure would be the kind of shock that forces Trump to choose between domestic political pain at the pump and a much larger ground presence in the region — a choice he has so far avoided by framing this as a blockade problem, not a war.
The bottom line
The June truce was always a pause, not a peace. The question was never whether the war would resume, but what would reignite it. The answer turns out to be the same thing the war was fought over in the first place: who controls the Strait of Hormuz, and on whose terms the world's oil flows.
Until that question is answered — diplomatically, militarily, or by exhaustion — every tanker that moves through those 21 miles of water is a bet on which side blinks first. Right now, neither is blinking. The market is starting to notice.
Sources
[1] AP News — "US expands strikes into northern Iran and disables ship trying to run blockade" (Jon Gambrell, July 16, 2026) — https://apnews.com/article/iran-us-hormuz-strait-war-july-16-2026-f98ff56554de2336f0e85bb5fdcae769
[2] BBC News — "Iran targets military bases as US launches wave of strikes" (Ella Kipling, July 16, 2026) — https://www.bbc.co.uk/news/articles/c2lq1ed28jxo
[3] BBC News — "Trump threatens to bomb bridges and power plants in Iran unless talks resume" (Jaroslav Lukiv & Helen Sullivan, July 2026) — https://www.bbc.com/news/articles/cy0608wy8pro
[4] Channel News Asia / Reuters — "Oil eases as traders lock in gains, assess Iran escalation" (July 16, 2026) — https://www.channelnewsasia.com/business/oil-prices-rise-4th-day-us-strikes-iran-raise-fears-wider-conflict-6257256