Inkling Is the First Western Open-Weight Model at Scale

Every open-weight model this publication has covered in 2026 has led with a number. GLM 5.2 beating GPT-5.5 on SWE-bench Pro. DeepSeek V4 leading LiveCodeBench globally. MiniMax M3 topping BrowseComp. Qwen 3.7 splitting benchmarks with Claude Opus. The entire frontier-scale open-weight story this year has been a story about Chinese labs -- GLM, DeepSeek, MiniMax, Qwen, Kimi, Ling -- racing each other to the top of leaderboards, at prices that made American closed models look expensive by comparison, while Google's Gemma line kept shipping smaller, genuinely well-built models that nobody mistook for frontier contenders and the rest of the American labs mostly stayed closed.

Mira Murati's Thinking Machines released its first model. It does not lead with a number. The company said so itself, in its own release materials: it is not the strongest model available today, closed or open.

That sentence is the story. Not because modesty is refreshing, but because it signals a genuinely different theory of what an open-weight model is for, and, as the correction below explains, it is the first time in this cycle that a serious Western lab has entered the open-weight fight at frontier scale on those terms, rather than staying closed, chasing the Chinese labs at their own benchmark game, or shipping something efficient but deliberately modest the way Gemma has for two years.

What Actually Shipped

Inkling is a mixture-of-experts model with 975 billion total parameters, activating roughly 41 billion for any given task, the same efficiency pattern this publication has covered in DeepSeek V3 and GLM 5.2, where a small fraction of the total parameter count does the work for any single query. It was trained on 45 trillion tokens spanning text, image, audio, and video, and reasons natively across all four modalities rather than routing them through separate specialist components. It is released under Apache 2.0, the most permissive license available, with full weights on Hugging Face including an NVFP4 checkpoint optimized for NVIDIA's newest Blackwell hardware.

Availability at launch is unusually broad for a first release. Inkling is live through APIs on Together, Fireworks, Modal, Databricks, and Baseten simultaneously, with inference partnerships already established for SGLang, vLLM, TokenSpeed, and llama.cpp. A model that ships with five API providers and four inference framework partnerships on day one is a model whose distribution strategy was planned as carefully as its training run.

The model card is explicit about intended use: general-purpose agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation, across English and other languages and multiple coding languages. That is a deliberately broad target. Inkling is not positioned as a coding specialist the way GLM 5.2 is, or a multimodal browsing specialist the way MiniMax M3 is. It is positioned as a competent generalist, the model you reach for across a wide range of tasks, rather than the model you reach for because it leads one specific leaderboard.

Why Not Winning Is the Actual Strategy

Thinking Machines raised a record $2 billion seed round at a$12 billion valuation in 2025, before releasing a single model or product. That is an extraordinary amount of capital to deploy on faith in a founding team, and it created real pressure to justify the valuation with a headline-grabbing launch. The obvious move, given that pressure, would have been to chase a benchmark win, train specifically toward SWE-bench Pro or Terminal-Bench 2.1, publish a chart showing a win over GPT-5.5 or Claude Opus 4.8, and let the number do the marketing, the way Grok 4.5's launch did days ago.

Thinking Machines did not do this. Inkling's own release materials state it performs well relative to similar open-weight offerings while explicitly declining to claim frontier status. The stated design goal is balancing cost against performance for broad, reliable use, not maximizing any single evaluation.

This is a real strategic bet, not a consolation prize dressed as modesty. A model tuned to win one specific benchmark is often a model that has been shaped, deliberately or not, around the particular shape of that benchmark's tasks, the exact failure mode this publication documented in the DeepSWE analysis, where Claude Opus 4.7 exploited SWE-bench Pro's git history and specific models showed dramatic score collapses when evaluated on a harder, less gameable benchmark. A model optimized for broad, decent performance across many task types is structurally less exposed to that kind of benchmark-specific overfitting, precisely because it was never optimized toward any single evaluation's particular blind spots.

The clearest demonstration of this generalist philosophy is Inkling fine-tuning itself. Using Tinker, Thinking Machines' existing model-adaptation product, Inkling autonomously wrote, ran, and evaluated its own fine-tuning job. That is not a benchmark score. It is a capability demonstration aimed squarely at the audience Inkling is actually built for: developers and enterprises who want a customizable foundation they can adapt to their own specific workload, rather than a frontier model they use exactly as shipped.

The Epistemics Bet

The most distinctive claim in Inkling's release is not about coding or agentic capability at all. It is about what Thinking Machines calls the model's epistemics, calibration, instruction following, and resistance to censorship, evaluated together as a single property.

Getting facts right requires more than memorizing a large training corpus. A model needs to express the right amount of confidence in its answers, including on questions that are not yet settled, a capability that matters directly for forecasting and prediction use cases, where fine-tuned models have recently begun outperforming general frontier LLMs. Thinking Machines evaluated Inkling on Cognition's Propaganda and Censorship Eval and reported that it exhibited strong patterns of censorship non-compliance, meaning the model answers directly on topics that other models, open or closed, tend to deflect or hedge on.

This is a specific and pointed differentiator against the open-weight competition Inkling is entering. Every major Chinese lab's open-weight release, DeepSeek, GLM, Qwen, MiniMax, Kimi, operates under a jurisdiction where the National Intelligence Law applies and where domestic political sensitivity around certain topics is a documented, well-understood constraint on model behavior. This publication has noted that footnote in every Chinese open-weight article this year: the models are genuinely capable and the license is genuinely permissive, but there are categories of question where the model's training almost certainly reflects domestic political constraints rather than an attempt at neutral, calibrated answering.

Inkling's censorship-resistance claim is Thinking Machines building a model specifically designed not to carry that particular constraint, while simultaneously reporting that it evaluated Inkling on genuinely dangerous-capability categories including CBRN, cyber risk, and loss of control, using both internal evaluation and external safety testers, and that it shows the strongest built-in safeguards of any open-weights model the company compared on FORTRESS, the benchmark that specifically tests refusal of weapons and violence requests alongside benign look-alike queries designed to trigger false refusals.

That combination, resist political censorship, maintain safety refusals on genuinely dangerous requests, is a harder needle to thread than either goal alone, and it is the first time an open-weight lab has framed its release specifically around that distinction rather than around raw capability.

Murati's Own History With This Exact Tension

There is a specific detail in Thinking Machines' positioning that deserves its own attention, because it comes from Murati's own past, not from marketing copy.

Murati was at OpenAI in 2019 when the lab, founded on an explicit promise of openness, withheld the full version of GPT-2 over misuse concerns, a decision that is now understood as the moment OpenAI began its retreat from fully open model releases. She was inside the organization that made that call, at the moment it made it.

Her own framing of Inkling's release suggests a case-by-case philosophy rather than an absolute commitment to openness: release models openly when the risks are manageable, and hold them back when they are not. Just because Inkling shipped open-weight does not mean every future Thinking Machines model will. That is a materially different starting position than DeepSeek's, GLM's, or MiniMax's consistent open-source strategy across their entire model lineups, and it is a materially different position than Anthropic's Fable-and-Mythos split, where the same underlying training produces one broadly available model and one tightly gated one.

Thinking Machines appears to be reserving that same optionality for itself from its very first release, having watched the GPT-2 decision from the inside and, by her own account, learned something from it about where the actual line sits.

Correcting the Record on "First"

One correction belongs here before going further, because getting it wrong would undercut the actual point. Inkling is not the first Western open-weight model. Google's Gemma line has been shipping open weights since 2024, moved to a permissive Apache 2.0 license with Gemma 4 in April, and remains, by volume of releases, the most consistent American open-weight effort running.

The honest assessment of Gemma is the reason Inkling's entry still matters. Nathan Lambert, one of the most closely followed independent AI researchers tracking open-weight releases, wrote plainly that the United States has already fallen behind in open models, both in performance and adoption rate, and pointed specifically at Gemma as the reason. The largest Gemma model tops out around 27 to 31 billion parameters: efficient, genuinely well-engineered for single-card deployment, and explicitly not a frontier-scale model built to lead on the coding and agentic benchmarks that actually drive enterprise adoption decisions. Google's own comparison chart at Gemma 4's April launch showed it landing slightly behind Qwen 3.5, GLM-5, and Kimi K2.5, not by a wide margin, but consistently behind, on a chart Google itself published.

The adoption data is more damning than the benchmark gap. The ATOM project, which tracks how many new open models worldwide are built as derivatives of which base model, found that from late 2023 through March 2026, roughly 70% of newly created global derivative models were built on top of Qwen. Llama's share, which stood near 40% two years earlier, had fallen to about 10%. Gemma does not register as a meaningful base for the derivative ecosystem at all in that accounting. Google has been shipping genuinely well-built small and mid-size open models for two years, and the global open-weight developer community has, in aggregate, kept building on Qwen instead.

This is the gap Inkling is actually entering, and it reframes what "first" means here. Inkling is not the first Western lab to release open weights. It is the first Western lab in this cycle to enter with the explicit ambition and scale to compete at the frontier tier that Gemma has never targeted, 975 billion total parameters against Gemma's 31 billion, a distribution strategy spanning five API providers on day one against Gemma's more modest AI Studio and Hugging Face footprint, and a positioning statement about broad reliability rather than edge efficiency. Gemma solved for accessible and small. Inkling is the first American attempt since the Chinese labs pulled ahead to solve for broad and large at the same time, under a license just as permissive as the one Google eventually adopted.

Whether Inkling succeeds where Gemma's scale ambitions never quite materialized is the open question. That it is even trying, at this parameter count, with this distribution plan, is itself the news.

Where This Leaves the Open-Weight Race

Before Inkling, the open-weight frontier conversation this publication has tracked all year was functionally a three-way fight between Chinese labs, with Gemma present but never contesting the top tier, the GLM 5.2, MiniMax M3, and DeepSeek V4 comparison piece from three weeks ago found all three Chinese labs sitting at the top of the Artificial Analysis Intelligence Index for open-weight models, with no Western lab anywhere near the top three, Gemma included.

Inkling does not immediately change that leaderboard. Thinking Machines said so itself. What it changes is the composition of who is competing at frontier scale at all. A model backed by $2 billion in seed funding, built by a former OpenAI CTO's team, shipped with day-one availability across five API providers and four inference frameworks, explicitly targeting broad reliability and customization rather than benchmark supremacy, is a genuinely different kind of entrant than anything Gemma has produced in two years or anything the Chinese open-weight cluster has produced this year, not because it is better than either, but because it is the first American attempt since Qwen and DeepSeek pulled ahead to compete at their scale rather than beneath it.

If the theory behind Inkling is correct, that a broadly competent, highly customizable, well-calibrated generalist has more durable value than a model tuned to win this month's specific benchmark, the real test will not be visible in next week's leaderboard. It will be visible in a year, in how many production systems were quietly built on Inkling as a foundation for fine-tuning, versus how many were built on whichever Chinese open-weight model held the top SWE-bench Pro score the week a team made its decision.

The Chinese labs are still winning the benchmarks. Whether they are still winning the actual adoption race in twelve months is now, for the first time in 2026, a genuinely open question.

Sources:

• Inkling: Our open-weights model -- Thinking Machines Lab
• Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling -- TechCrunch
• Mira Murati's Thinking Machines debuts its first AI model -- Axios

Previously on TheQuery: The Real AI Race Nobody Is Covering: MiniMax M3, GLM 5.2, and DeepSeek V4 Are Fighting Each Other and Grok 4.5 Is Not "Opus-Class." It Is Something More Interesting: Opus-Adjacent and Much Cheaper.