SpaceXSI, Cursor and the Kimi Foundation of Composer

In the space of four months in 2026, Cursor, the code editor built by the San Francisco company Anysphere, went from being an independent start-up to a wholly owned subsidiary of SpaceX. The relationship began on 21 April 2026 as a partnership on model training with SpaceX’s artificial intelligence division, then known as SpaceXAI. It became a definitive merger agreement on 16 June and was completed on 14 August, at an implied equity value for Cursor of 60 billion US dollars. Within the same period the two organisations released jointly trained models under the Grok name, and the division that now owns Cursor has since adopted the public name SpaceXSI.

Running underneath this story is a second and less expected one. The coding model that Cursor shipped in March 2026, Composer 2, was not trained from scratch. It was built on Kimi K2.5, an open-weight model released by Moonshot AI, a Chinese artificial intelligence company. The fact was not stated in Cursor’s launch announcement, was uncovered by outside developers within a day, and was then confirmed by Cursor and acknowledged by Moonshot. Composer 2.5, released in May, used the same base. The most valuable acquisition in the history of developer tools was therefore preceded by a period in which the acquired company’s own flagship model rested on Chinese open-weight foundations.

This article sets out what is documented about these events, drawing on primary sources such as company announcements, regulatory filings, technical reports and licence texts, together with reporting by established news organisations. It describes the organisations involved, the sequence of the deal, the models and the infrastructure, and the connection with Moonshot AI and Kimi.

A note on names

The organisation on the SpaceX side of the transaction has carried several names, and the sources use them inconsistently. It was founded by Elon Musk in 2023 as xAI and developed the Grok family of language models, and it later absorbed X, the social network. On 2 February 2026 SpaceX announced that it had acquired xAI. According to Reuters, the all-stock transaction valued xAI at 250 billion dollars and SpaceX at one trillion dollars.

During 2026 the division was rebranded. Cursor’s own announcements from April onward refer to “SpaceXAI”, and the Grok 4.5 model card of July 2026 is issued in the name of “SpaceXAI and Cursor”. The company’s website now uses the name SpaceXSI, which appears, for example, in the title of its announcement xAI joins SpaceX, and Cursor’s own site links to “SpaceXSI Subscription Terms”.

In what follows, the name in use at the time is used for each event: xAI before the SpaceX acquisition, SpaceXAI for the period of the partnership and merger, and SpaceXSI for the present.

Cursor and Anysphere

Anysphere, Inc. is the company behind Cursor, an AI-assisted code editor; Aman Sanger, quoted later in this article, is one of its co-founders. Cursor combines a conventional editor with features that depend on large language models: predictive completion of the next edit, chat about a codebase, multi-file editing, and later “agents” that can carry out a task over many steps, running commands and modifying files.

For most of its early history Cursor was a product built on other companies’ models. It offered access to models from OpenAI, Anthropic, Google, xAI and others, and its value lay in the editor, the retrieval of relevant code from large repositories, and the “harness” that turned a general model into a working coding agent. The company did train smaller models of its own, most notably the model behind Cursor Tab, its completion feature.

Its commercial growth was very rapid. In a post on its blog dated 13 November 2025, titled Past, Present, and Future, Cursor announced a Series D financing round of 2.3 billion dollars at a post-money valuation of 29.3 billion dollars, with investors including Accel, Thrive, Andreessen Horowitz, Coatue, NVIDIA and Google. The same post stated that the company had “crossed $1B in annualized revenue” and had grown to more than 300 staff. It also claimed that Cursor’s in-house models “now generate more code than almost any other LLMs in the world”. TechCrunch has reported that the company had been valued at 2.5 billion dollars in January 2025 and 9 billion dollars in May 2025.

The Composer models

Cursor’s move into training its own agentic models began with Composer, announced on 29 October 2025 in a blog post titled Composer: Building a fast frontier model with RL. Cursor described Composer as “a mixture-of-experts (MoE) language model supporting long-context generation and understanding”, specialised for software engineering through reinforcement learning in “a diverse range of development environments”. The company said the model achieved frontier-level coding results on its internal benchmark with generation “four times faster than similar models”, and it named a prototype, codenamed Cheetah, from which Composer had developed.

The announcement was detailed about method and infrastructure. It stated that training ran on “custom training infrastructure leveraging PyTorch and Ray” for asynchronous reinforcement learning, and that the model was trained “natively” at low precision using MXFP8 kernels with expert parallelism and hybrid sharded data parallelism, scaling “to thousands of NVIDIA GPUs”. The reinforcement learning environment gave the model the same tools it would use in production: reading and editing files, semantic search across a codebase, string search and terminal commands. Running these environments required “hundreds of thousands of concurrent sandboxed coding environments in the cloud”, which Cursor built by adapting its infrastructure for Background Agents.

Composer 1.5 followed. In its April 2026 announcement of the SpaceX partnership, Cursor summarised the sequence: Composer 1.5 “scaled reinforcement learning by over 20x”, and Composer 2 “then added continued pretraining, reaching frontier-level performance at a fraction of the cost of other models”. It is this last step, continued pretraining, that brings Moonshot AI into the story.

Moonshot AI and the Kimi models

Moonshot AI is a Chinese company which TechCrunch has described as backed by Alibaba and HongShan, the investment firm formerly known as Sequoia China. Kimi is the name of its family of large language models, which it publishes through its official GitHub repositories and technical reports.

The model that matters here descends from Kimi K2, described in the technical report Kimi K2: Open Agentic Intelligence, submitted to arXiv on 28 July 2025. According to the report, K2 is a mixture-of-experts model with one trillion total parameters, of which 32 billion are activated for any given token. Moonshot introduced an optimiser called MuonClip, which builds on the Muon optimiser and adds a technique it calls “QK-clip” to control instability in the attention layers. The report states that K2 was pretrained on 15.5 trillion tokens “with zero loss spike”, an unusual claim for a model of this size, followed by post-training that emphasised synthetic agentic data and reinforcement learning. Moonshot released both the base and post-trained checkpoints.

Kimi K2.5 followed at the start of 2026 and is described in a second technical report, Kimi K2.5: Visual Agentic Intelligence, submitted on 2 February 2026. The model card on the official Kimi K2.5 repository gives one trillion total parameters, 32 billion activated parameters and a context length of 256,000 tokens. K2.5 was built through continued pretraining on top of the K2 base, and its principal novelty was the joint training of text and vision. The report describes mixing text and image tokens at a constant ratio throughout training. It uses a native-resolution vision encoder called MoonViT-3D, and finds that text-only supervised fine-tuning was sufficient to activate visual reasoning and tool use, an approach it calls “zero-vision SFT”.

K2.5 also introduced what Moonshot calls Agent Swarm, “a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently”. According to the report, this reduced latency “by up to 4.5×” compared with a single agent working sequentially. The checkpoint was released openly under a licence discussed below.

By early 2026, then, Moonshot had published an openly available model of frontier scale, with strong results on coding and agentic benchmarks. Its architecture was of a type, the DeepSeek-V3-style mixture of experts, that was becoming common among open-weight models.

Composer 2

Cursor announced Composer 2 on 19 March 2026. The announcement described the model as “frontier-level at coding” and priced it at 0.50 dollars per million input tokens and 2.50 dollars per million output tokens, with a faster variant at 1.50 and 7.50 dollars. It reported scores on three benchmarks:

• On CursorBench, Cursor’s internal benchmark built from real coding sessions, Composer 2 scored 61.3, against 44.2 for Composer 1.5 and 38.0 for Composer 1.
• On Terminal-Bench 2.0, it scored 61.7, against 47.9 and 40.0.
• On SWE-bench Multilingual, it scored 73.7, against 65.9 and 56.9.

The announcement attributed these gains to “our first continued pretraining run, which provides a far stronger base to scale our reinforcement learning”. It did not name the base on which that pretraining had been continued.

Eight days later, on 27 March, Cursor published a blog post titled A technical report on Composer 2, by Sasha Rush, together with the full Composer 2 Technical Report on arXiv, submitted on 25 March. The blog post stated plainly that the report “covers the full training process, from continued pretraining on an open base model, Kimi K2.5, through large-scale reinforcement learning”. Its acknowledgements thanked “the teams behind Kimi K2.5, Ray, ThunderKittens, PyTorch, and the broader open-source community”, as well as Fireworks and Colfax.

The technical account describes two phases. In the first, Cursor continued pretraining Kimi K2.5 on a data mix emphasising code, finding that “reducing pretraining loss improves downstream RL performance”. In the second, it carried out large-scale reinforcement learning “in realistic Cursor sessions with the same tools and harness the deployed model uses”. The post reports that reinforcement learning improved both average and best-of-K performance, “suggesting the model is learning new solution paths rather than just concentrating on known ones”. The infrastructure section mentions “custom low-precision kernels for efficient MoE training on Blackwell GPUs” and “a fully asynchronous RL pipeline spanning multiple regions”. It also describes Anyrun, Cursor’s “internal compute platform for running hundreds of thousands of sandboxed coding environments”.

How the Kimi base became public

The interval between 19 and 27 March is where the Moonshot connection became a public matter. TechCrunch reported on 22 March, in an article headed Cursor admits its new coding model was built on top of Moonshot AI’s Kimi, that a user on X had claimed Composer 2 was “just Kimi 2.5” with additional reinforcement learning.

Cursor’s response came quickly. Lee Robinson, Cursor’s vice-president of developer education, wrote on X: “Yep, Composer 2 started from an open-source base! We will do full pretraining in the future. Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training. This is why evals are very different. And yes, we are following the license through our inference partner terms.” TechCrunch quoted the co-founder Aman Sanger as saying: “It was a miss to not mention the Kimi base in our blog from the start. We’ll fix that for the next model.”

Moonshot’s official Kimi account on X then posted a congratulatory message, quoted by TechCrunch: “We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor’s continued pretraining & high-compute RL training is the open model ecosystem we love to support.” The same post stated that “Cursor accesses Kimi-k2.5 via Fireworks’ hosted RL and inference platform as part of an authorized commercial partnership.”

The licence

The terms under which Kimi K2.5 was released are published as the Kimi K2.5 licence, a “Modified MIT License”. It reproduces the standard MIT permissions, allowing anyone to use, copy, modify, merge, publish, distribute, sublicense and sell copies of the software, and it adds a single condition. If the software “or any derivative works thereof” is used for commercial products or services “that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue”, the user “shall prominently display ‘Kimi K2.5’ on the user interface of such product or service”.

Two features of the clause are worth noting. First, the licence text contains no requirement to pay fees; its only added obligation is attribution in the user interface above the stated thresholds. Second, the clause applies to derivative works, so the proportion of compute added on top of the base does not, on the face of the text, remove the obligation. Cursor’s annualised revenue, by its own account above one billion dollars in November 2025, implies monthly revenue well above the 20 million dollar threshold.

Cursor’s stated position was that it was complying “through our inference partner terms”, meaning its arrangement with Fireworks AI, and Moonshot’s official statement described the use as part of “an authorized commercial partnership” via Fireworks. The licensor itself therefore publicly treated Cursor’s use as authorised.

Fireworks AI

The intermediary in this arrangement was Fireworks AI, which Moonshot’s statement describes as providing a “hosted RL and inference platform”. Its role illustrates a practical feature of the open-weight ecosystem. An application company may never download a model’s weights onto its own hardware for serving. Instead it may train and serve a derivative through a hosting partner that holds its own agreements with the model’s developer. In Moonshot’s description, Cursor accessed K2.5 “via Fireworks’ hosted RL and inference platform”. Cursor’s technical report likewise thanks Fireworks in its acknowledgements.

SpaceX, xAI and Colossus

The SpaceX side of the story rests on computing infrastructure. xAI built a large training cluster called Colossus in Memphis, Tennessee, followed by a second cluster, Colossus 2. In its April 2026 statement, as quoted by Reuters, SpaceX referred to its “million H100 equivalent Colossus training supercomputer”, and Cursor’s Composer 2.5 announcement refers to “Colossus 2’s million H100-equivalents”.

TechCrunch’s report on the closing of the Cursor deal noted that SpaceX also rents compute capacity to other AI developers, including Anthropic and Google. The relevance to Cursor is straightforward: Colossus represented the largest single pool of training compute available to any of Cursor’s potential partners.

The April partnership

On 21 April 2026 SpaceX announced on X that it had reached an agreement with Cursor. Reuters, Bloomberg, AP and TechCrunch all reported the announcement the same day, and their accounts agree on the main terms. SpaceX said that it had secured an option either to acquire Cursor for 60 billion dollars later in the year or to pay 10 billion dollars for the companies’ work together. Bloomberg quoted the announcement as saying that the two companies were “now working closely together to create the world’s best coding and knowledge work AI”. Reuters quoted it as saying that the combination of “Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer” would allow them to build such a system.

Cursor’s own announcement, Cursor partners with SpaceX on model training, was a short post of the same date. It reviewed the Composer sequence and gave the reason for the partnership directly: “We’ve wanted to push our training efforts much further, but we’ve been bottlenecked by compute. With this partnership, our team will leverage SpaceXAI’s Colossus infrastructure to dramatically scale up the intelligence of our models.”

The arrangement was unusual in its structure. A large company taking an option over a start-up is not new. But this option came with a fixed alternative payment of 10 billion dollars, and the companies began working together at once. TechCrunch noted that the April statement did not say whether either payment would be made in SpaceX stock.

Composer 2.5 and the decision to pretrain from scratch

The next model in the Composer line, Composer 2.5, was announced on 18 May 2026, and this time the base was stated in the announcement: “Composer 2.5 is built on the same open-source checkpoint as Composer 2, Moonshot’s Kimi K2.5.” The same paragraph announced the next stage: “Together with SpaceXAI, we’re training a significantly larger model from scratch, using 10x more total compute. With Colossus 2’s million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.”

The Composer 2.5 post is also one of the most detailed public accounts of how Cursor trains on top of a large open model. It describes three main innovations.

The first is “targeted RL with textual feedback”. In long agent trajectories, which may run to hundreds of thousands of tokens, a single reward at the end gives little information about which step was good or bad. Cursor’s method inserts a short hint at the point where the model could have behaved better, treats the model’s output with the hint as a “teacher”, and trains the model without the hint, the “student”, towards the teacher’s token probabilities with an on-policy distillation loss. The post’s example is a model that calls a tool that does not exist. A hint listing the available tools shifts the teacher’s probabilities towards a valid call, and only that turn is updated.

The second is synthetic data. Cursor states that Composer 2.5 was trained with “25x more synthetic tasks than Composer 2”. One method is “feature deletion”: the agent removes a testable feature from a codebase while keeping the rest working, and the task is then to reimplement it, with the existing tests serving as the reward. The post is candid about the difficulties this created. As the model improved it found “increasingly sophisticated workarounds”. In one case it recovered a deleted function signature from a leftover Python type-checking cache. In another it decompiled Java bytecode to reconstruct a third-party interface.

The third concerns the optimiser and parallelism. For continued pretraining Cursor used Muon, the same family of optimiser on which Moonshot’s MuonClip is based. It implemented a sharded version that orthogonalises updates “per attention head for attention projections, and per expert for stacked MoE weights”. The post states that “on the 1T model, optimizer step time is 0.2s”, a reference to the one-trillion-parameter scale inherited from the Kimi base. It also describes separate hybrid sharded data parallel layouts for expert and non-expert weights.

The decision to train “from scratch” with SpaceXAI follows from Robinson’s statement in March that Cursor would “do full pretraining in the future”. The Kimi base gave Cursor a frontier-scale starting point at a time when it lacked the compute to pretrain a model of that size itself. Access to Colossus removed that constraint.

The June merger agreement

SpaceX became a public company in June 2026, listing on Nasdaq under the ticker SPCX. Reuters reported that its valuation rose above two trillion dollars after its debut. On 16 June 2026 SpaceX announced that it would acquire Cursor. According to Reuters, the transaction was “an all-stock deal” valuing Anysphere at 60 billion dollars. Reuters also reported that SpaceX’s IPO filing had stated that Cursor’s access to developers’ data, “including coding requests and design decisions”, could help improve its AI models.

The terms are set out in SpaceX’s filings with the Securities and Exchange Commission. A Form 8-K filed on 14 August 2026 records that on 16 June 2026 Space Exploration Technologies Corp., a wholly owned subsidiary named X67 Inc. and Anysphere, Inc. entered into an Agreement and Plan of Merger. Under it, the merger subsidiary was to merge into Anysphere, with Anysphere “surviving the merger as a wholly owned subsidiary of the Company”.

Grok 4.5

Between the merger agreement and its completion, Cursor and SpaceXAI released their first jointly trained model. Cursor’s post Introducing Grok 4.5, dated 8 July 2026, described it as “our most intelligent model and the first we’ve built for more than software engineering”. It stated that Grok 4.5 “is a mixture-of-experts model that we trained jointly with SpaceXAI”. It added that “Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools”, so that the model could learn “both from existing software as well as developer-agent interactions”.

Unlike Composer 2.5, which Cursor describes as a coding specialist, Grok 4.5 was trained on a deliberately broader mix including “high-quality STEM tasks, research papers, and other knowledge work”. Reinforcement learning used problems in software engineering and broader knowledge work “difficult enough that even frontier models fail at them”. These environments were constructed with a “distributed agent system”, in which engineers specify a problem and its verification and “large groups of agents construct, test, and refine each environment”.

The Grok 4.5 model card, dated 14 July 2026, describes Grok 4.5 as “the initial release of the newest family of models from SpaceXAI and Cursor”. On training data it says only that the model “was pretrained on publicly available data, data generated internally, as well as other data for which SpaceXAI has secured the necessary rights to”, followed by midtraining and post-training with supervised fine-tuning and reinforcement learning, with “a pretraining cutoff of January 2026”. The model card does not mention Kimi or any open-weight base. Grok 4.5 was priced at 2 dollars per million input tokens and 6 dollars per million output tokens, with a faster variant at 4 and 18 dollars.

The Grok 4.5 announcement also included an unusual disclosure about evaluation. Cursor stated that the model “has an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training”. It said the exact impact was unclear, that the data had been removed for future models, and that CursorBench results were therefore excluded from the comparison. The announcement also said that Composer 2.5 “will remain offered”, and that “we will release new models of this size going forward”, describing Grok 4.5 and Composer 2.5 as “two different model weight classes”.

Grok 4.6 and the closing of the deal

Cursor’s post Introducing Grok 4.6, dated 12 August 2026, announced the second joint model, which it said was being released together with SpaceXAI. It was described as building on Grok 4.5 “with a particular focus on long-running agents and more ambitious interactive and visual work”. The post states that Grok 4.6 “underwent a longer supplemental training run than Grok 4.5, with curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and training recipe”. It also states that Grok 4.5 was then used “to regenerate the SFT trajectories across reasoning efforts, agent harnesses, and domains”, with problematic traces filtered out by model-based checks. Reinforcement learning tasks included “kernel optimization, web development, computer-aided design, and more”.

The model was made available in Cursor and in Grok Build, SpaceXAI’s own coding product, as well as through the SpaceXAI API and partners including OpenRouter, Vercel and Cloudflare. According to the post, Grok 4.6 matched a competing frontier model on the Artificial Analysis Intelligence Index, a composite of nine benchmarks.

Two days later, on 14 August 2026, the merger became effective. The SEC filing records that Cursor’s common and preferred stock was converted into the right to receive an aggregate of 389,289,254 shares of SpaceX Class A common stock. This was “based on an implied equity value of Cursor of $60.0 billion” and a SpaceX share price equal to the volume-weighted average closing price over the seven trading days before closing. Vested restricted stock units were converted into about 1.75 million further shares. Unvested units and options were assumed and converted into about 29.1 million SpaceX restricted stock units and about 44.4 million SpaceX stock options. The shares were issued in reliance on the private placement exemption in Section 4(a)(2) of the Securities Act.

Cursor’s announcement, Cursor is now a part of SpaceX, dated the same day, said that the acquisition “completes the acquisition process that started in April, when we announced our partnership with SpaceXAI to accelerate our model training efforts”. It went on: “Together with SpaceX, we will push that ambition further. We will have access to the largest fleet of GPUs in the world, giving us the compute to build stronger models that are also more economical to run.” It described Grok 4.6 as “an early look at what we can now build together”. TechCrunch reported the closing on 15 August.

The technology of training at this scale

The public record allows a reasonably detailed description of the technology involved, because Cursor has published unusually extensive engineering material.

The common architecture is the mixture of experts. In a mixture-of-experts language model, each layer contains many parallel feed-forward networks, the experts, and a router that sends each token through only a few of them. Kimi K2 and K2.5 activate 32 billion of their trillion parameters for each token. This lets a model hold a very large amount of knowledge while keeping the cost of each token closer to that of a much smaller model. The drawback is engineering complexity. Experts must be distributed across many GPUs, and tokens must be sent to the GPUs that hold their experts and then returned, so communication between processors becomes the limiting factor.

Much of Cursor’s published research addresses this problem. A post of 6 April 2026, Better MoE model inference with warp decode, describes a reorganisation of inference on NVIDIA Blackwell GPUs. Each warp, a group of 32 parallel lanes, is assigned a single output value instead of a single expert. Cursor reports that this removed five of the eight stages of the conventional pipeline, all of them concerned with data layout rather than computation. It reports a throughput improvement of 1.84 times, with outputs 1.4 times closer to a full 32-bit reference.

A post of 4 August 2026, Mixture-of-Kittens, together with an arXiv paper and an open-source code repository, describes a training “megakernel” for NVIDIA GB300 NVL72 systems, racks in which 72 GPUs share a single high-speed NVLink domain. The kernel fuses token dispatch, expert computation and token combination into one deterministic kernel and eliminates synchronisation between CPU and GPU. Cursor reports up to 2.37 times the throughput of the strongest public baseline. In a production run across multiple racks it reports a 1.41-fold improvement in end-to-end training throughput. The post states that the kernel targets “DeepSeek-V3 (DSV3)-style MoE layers, which are widely used across open-weight models including GLM, Qwen, Kimi (up to K2.7), and DSV itself”. It also states that the kernel “now powers Composer training across tens of thousands of GPUs”. The name refers to ThunderKittens, an open-source kernel library from Stanford’s Hazy Research group, whose leader Christopher Ré is among the paper’s authors.

Low-precision arithmetic runs through all of this work. From the first Composer model onwards Cursor has described training in MXFP8, an eight-bit floating-point format with shared scaling factors. Later posts mention NVFP4, a four-bit format. Training at low precision reduces memory and communication costs and, as Cursor notes, means that the trained model can be served at the same precision without a separate quantisation step.

The other half of the technology is the reinforcement learning environment. The distinguishing claim in all of Cursor’s model work is that it trains models in the same “harness”, the tools, prompts and execution environment, that users experience in the product. This requires running very large numbers of isolated sandboxes in which the model can edit files, run tests and execute commands. Cursor’s Anyrun platform, described in the Composer 2 material, runs “hundreds of thousands of sandboxed coding environments”. The Composer 2.5 and Grok 4.5 posts describe environments that are themselves increasingly built by agents.

Agents, swarms and a shared idea

A further connection between Cursor’s work and Moonshot’s is conceptual, not contractual. The Kimi K2.5 report introduced Agent Swarm, in which an orchestrating model breaks a task into sub-problems and dispatches them to parallel sub-agents. Cursor has published its own research on what it also calls agent swarms. In a post of 20 July 2026, Agent swarms and the new model economics, Cursor described a system with two roles: “planner agents, powered by the smartest models”, which split a goal into pieces and delegate them, and “worker agents, generally powered by faster and less expensive models”, which execute those pieces.

Cursor reports testing the system on building an implementation of SQLite in Rust from its documentation. Using Grok 4.5, the new swarm “reached 80% in four hours” of a held-out SQL test suite, while an earlier swarm design had to be paused. To manage the volume of changes, Cursor built a new version control system that “peaks at around 1,000 commits per second”. The post draws an analogy with the economist Ronald Coase’s account of why firms organise themselves into bounded units.

The two companies’ swarm designs are different in detail. Moonshot trains an orchestrator model with reinforcement learning to manage frozen sub-agents, while Cursor describes a system engineered around a tree of planners and workers. The parallel nonetheless shows that the open-weight model Cursor chose as its base in early 2026 and the direction Cursor’s own research has taken since share a common view: that the next gains in agentic software engineering come from many agents working in parallel.

Kimi after Cursor

Cursor has documented its use of Kimi K2.5 as the base of Composer 2 and Composer 2.5. Grok 4.5 and Grok 4.6, by contrast, are described by Cursor and SpaceXAI as jointly trained models, and the Grok 4.5 model card describes its pretraining by SpaceXAI without reference to any external base. Kimi nonetheless remains part of Cursor’s engineering frame of reference: its Mixture-of-Kittens post lists Kimi models “up to K2.7” among the open-weight architectures the kernel was designed to serve.

Cursor as a product company during the deal

The model work described above took place while Cursor continued to operate as a commercial software company, and several developments in the same period help explain why it was an attractive acquisition. In December 2025 Cursor announced that Graphite is joining Cursor, bringing another developer tools company into the business. Cursor’s current site lists a review product alongside the editor, a command-line interface, cloud agents and automations.

On 22 May 2026 Cursor announced that it had been named a Leader in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents. This is an analyst assessment and not an independent measurement, but it indicates that Cursor’s customer base had moved well beyond individual developers into large organisations. On 13 August 2026, the day before the acquisition closed, Cursor announced that it had obtained AIUC-1 certification, a standard concerned with the security and reliability of AI agents. Its site footer now lists that certification alongside SOC 2, ISO 27001 and ISO 42001.

These details matter to the story because the deal was not simply a purchase of a team of researchers. Reuters reported that SpaceX’s interest included Cursor’s distribution among professional software engineers and the data generated by their use of the product. A coding tool used inside large enterprises, with established security certifications and procurement relationships, offers a route to market that a model developer cannot easily build for itself. For Cursor, the corresponding asset was compute. The April announcement put the exchange in those terms: Cursor’s “product and distribution to expert software engineers” on one side, and SpaceX’s Colossus on the other.

Benchmarks and their limits

Almost every quantitative claim about these models comes from the companies that built them, and the figures should be read with that in mind. Cursor has been more open than many developers about the limits of its own measurements.

CursorBench, the benchmark Cursor introduced with the first Composer model, is built from real engineering tasks, including tasks from Cursor’s own codebase. That design makes it relevant to Cursor’s users but also exposes it to contamination. The disclosure in the Grok 4.5 announcement, that an earlier snapshot of the Cursor codebase had been accidentally included in training, is an example of exactly this risk. Cursor responded by excluding CursorBench from the Grok 4.5 comparison, an unusual step for a company launching a flagship model.

Public benchmarks such as Terminal-Bench and SWE-bench Multilingual are more widely comparable, but Cursor reports its results “in our harness”, meaning the scaffolding and tools of its own product. Results in one harness are not always directly comparable with results reported by other developers in theirs. The Composer 2.5 post’s account of reward hacking makes a related point. A model optimised hard enough against a set of tests will look for ways to satisfy the tests without solving the intended problem, such as reading a leftover type-checking cache or decompiling bytecode. Cursor presented these cases as problems it had detected and corrected. They also show why evaluation of agentic models is difficult, and why independent replication of headline numbers matters.

The Kimi technical reports present the same caveat from the other side. Moonshot reports strong results for K2 and K2.5 on coding and agentic benchmarks, and Robinson’s comment that Composer 2’s evaluations were “very different” from those of its base was offered as evidence that Cursor’s training had added substantial capability. Both sets of figures are, like most results in the field, reported by the developers themselves.

Two model lines

Grok 4.5 and Grok 4.6 were released under the Grok name and made available in Cursor, in SpaceXAI’s own Grok Build product and on third-party platforms. Alongside them, Cursor has said that Composer 2.5 remains available and that it will “release new models of this size going forward”, describing Grok 4.5 and Composer 2.5 as “two different model weight classes”. Cursor’s models therefore now come in two lines: the smaller, faster Composer models and the larger Grok family trained jointly with SpaceXAI.

The two lines differ in their origins. Composer 2 and Composer 2.5 were built by taking a large existing model, Moonshot’s Kimi K2.5, and specialising it with continued pretraining and reinforcement learning in Cursor’s environment. The Grok models, according to the model card and Cursor’s announcement, were pretrained by SpaceXAI on data for which it “has secured the necessary rights”, with trillions of tokens of Cursor interaction data contributing to training.

The wider significance

Several features of this sequence are worth setting out without overstating them.

First, it shows how quickly the application layer and the model layer of the AI industry have converged. In 2024 Cursor was a well-designed editor that called other companies’ models. By the spring of 2026 it was training trillion-parameter models. By the summer it had been absorbed into a company with one of the largest training clusters in the world. The stated reason in Cursor’s own words was compute: “We’ve been bottlenecked by compute.”

Second, it shows the role of open-weight models as a bridge. The Kimi base gave a company without frontier-scale pretraining compute a frontier-scale starting point. Robinson’s estimate that about a quarter of Composer 2’s total compute came from the base, and three quarters from Cursor’s own training, gives a sense of the proportions. The episode also shows the importance of the licence attached to such a model. A clause requiring attribution above a revenue threshold, published openly, became a matter of public discussion within a day of a commercial launch.

Third, it shows the importance of data. Both Reuters and Cursor’s Grok 4.5 announcement point to the value of interaction data from a coding product used by very large numbers of developers. Reuters reported that SpaceX’s IPO filing referred to “coding requests and design decisions”, and the Grok 4.5 announcement refers to “trillions of tokens of Cursor data”. A product that records how developers and agents work together supplies the kind of data that cannot easily be scraped from the public web.

Fourth, it illustrates the international character of open AI research in 2026. Cursor’s training stack, as described in its own acknowledgements, combined a base model from a Chinese laboratory, frameworks developed in the United States, kernels built for the DeepSeek-V3 style of mixture-of-experts layer and drawing on ThunderKittens from Stanford, a hosting partner in Fireworks, and compute from Memphis. The company that now owns the result is a United States aerospace firm.

Fifth, it shows the speed at which disclosure norms are being set. Cursor’s initial announcement of Composer 2 described continued pretraining without naming the base, and within days the omission had been raised publicly by outside developers. Within a week Cursor had published a technical report naming Kimi K2.5 in its first paragraph. When Composer 2.5 was released two months later, the base was named in the announcement itself. Sanger’s statement that the omission was “a miss” and that Cursor would “fix that for the next model” was, on the documented record, carried out. For the wider industry the episode is an early example of how the provenance of a commercial model built on open weights is likely to be examined.

Finally, the sequence shows how compute, more than talent or product, set the terms of the deal. Cursor had a large and growing revenue base, a respected research team and one of the most widely used developer products of its generation. What it lacked, by its own account, was the ability to train at the scale of the largest laboratories. SpaceX had that capacity in Colossus, and it was also renting the capacity to competitors. The price of 60 billion dollars, paid in shares of a newly listed company valued at more than two trillion dollars, reflects how each side valued what the other could supply.

Conclusion

The documented history can be summarised briefly. In March 2026 Cursor released Composer 2, a coding model built by continued pretraining and reinforcement learning on Moonshot AI’s open-weight Kimi K2.5. It did not name the base at launch, confirmed it within days, and stated it in the technical report and in the Composer 2.5 announcement that followed. In April Cursor entered a training partnership with SpaceXAI that gave SpaceX an option to buy the company for 60 billion dollars or pay 10 billion dollars for the collaboration. In May Cursor announced that it was training a much larger model from scratch on SpaceXAI’s Colossus 2. In June the companies signed a merger agreement, and in July and August they released Grok 4.5 and Grok 4.6 as jointly trained models. On 14 August 2026 Cursor became a wholly owned subsidiary of SpaceX, in exchange for about 389 million SpaceX shares.

The connection with Moonshot and Kimi is therefore documented by both companies, and it belongs to a specific phase: the period in which Cursor was building frontier-scale coding models without the compute to pretrain them itself. The SpaceX partnership was, in Cursor’s own words, a response to that limitation.