On Saturday 10 October the Financial Times reported that Nvidia is in early-stage talks to acquire Reflection AI or to increase its investment in the company. I could not read the FT piece itself, so what follows rests on the Investing.com write-up of it. That write-up says the people familiar with the matter described a possible full acquisition, more funding, or an acqui-hire, meaning Nvidia would hire Reflection’s staff and license its technology. It says a deal could come within weeks, that the talks are preliminary and could collapse, and that neither company had immediately responded to requests for comment. Reuters, per the same write-up, could not independently verify the report. So the only fact on the table is that the FT said something. Anyone building on it should know that.
The timing matters. Five days earlier, on 5 October, Reflection announced Beam, “Reflection’s first open-weight model.” This essay is about the gap between those two words, open and weight, and the file that is supposed to close it.
What Beam is, and is not, on 10 October
The announcement gives specifications. Beam has 501 billion total parameters with 23 billion active, was pretrained on 23.8 trillion tokens on a cluster of 6,144 Nvidia GB300 NVL72 GPUs, and then went through four weeks of reinforcement learning on 10.5 thousand GB300 GPUs. It lists scores of 80.9 on SWE-Bench Verified, 80.1 on Terminal Bench v2.1, 97.8 on AIME 2026, 90.5 on GPQA Diamond and 36.2 on Humanity’s Last Exam without tools. Those are the company’s tables, updated on 8 October when the post says the results were revised, and the post says the other models’ evaluation data in its efficiency comparison comes from Artificial Analysis and DataCurve, and that its FLOPs figures are estimates, not measured inference costs. I have no independent run of any of them, and nobody outside Reflection can have one yet, because nobody outside a waitlist can run the model.
That is the central fact. The post says “We will release the weights, technical report, model card, and developer artifacts later this month,” and that the weights will be under an Apache 2.0 license. Early access is limited to “a select group of users,” with a waitlist. When I checked Reflection’s Hugging Face organization on 10 October, its Models and Datasets sections both read “None public yet.” There is no license file to read, so I cannot tell you what the license says. I can tell you what the blog says it will say.
I have written about this pattern before. Mistral Large 4’s preview was described as an open model while the post promised that the weights would “drop end of this month.” The word is not wrong in either case; it is early. A model is open-weight on the day the weights can be downloaded, and the useful thing a reader can do before that day is write down exactly what has been promised, so that the day can be graded.
What “open” would have to include
The word has a reference definition. The Open Source Initiative’s Open Source AI Definition, version 1.0, asks for four freedoms (use, study, modify, share) and says they require access to the “preferred form to make modifications.” It breaks that into three parts: sufficiently detailed information about the training data, the complete source code used to train and run the system, and the parameters, meaning the weights. The data itself does not have to be released; the information about it does.
Set that against what Reflection has said. Apache 2.0 weights, if they arrive as promised, would cover the third part. The Beam post describes its data in general terms (pretraining tokens “from the web, public sources, and proprietary licensed datasets,” with “proprietary vendor data” among the sources for its reinforcement-learning environments), but it promises neither training code nor training-data information. In October 2025, TechCrunch reported that Reflection planned to release weights while “largely keeping datasets and full training pipelines proprietary.” The same article said the company had built a training stack it promised would be “open for all.” Those two statements sit awkwardly together, and I found no later statement that resolves it.
In fairness to Reflection, its own wording is narrower than the definition. The post says open-weight, not open-source, and the narrower word is the accurate one. If the files ship as described, “open-weight” will be a fair description. “Open source AI” in the Initiative’s sense is, on the present record, unestablished, which is a different claim from false. The claim that is falsifiable now is simpler: that Beam can be downloaded. It cannot.
Why a deal makes the file matter more
Suppose the talks go somewhere. The FT’s three structures differ in what they leave behind. A full acquisition puts the team, the weights and any unreleased successors inside Nvidia. More funding changes little visible. An acqui-hire, as the write-up describes it, moves people and licenses technology, and, according to the same account, could avoid the lengthy regulatory review a takeover would invite. I cannot judge which is likelier, and the report says only that the talks are preliminary.
What I can say is what a release changes. The Apache 2.0 text, as I read it, grants a perpetual and irrevocable license to the version you receive, subject to its conditions (for example, the patent license ends for a licensee who sues over patents in the work). Beam’s actual file does not exist yet, so this is a reading of the standard terms, not of Reflection’s. If the weights are published before any deal closes, those copies stay usable by anyone who downloaded them, whoever owns the company next year. What no license in the file can cover is everything after it: the next model, the people who made it, the stack that is promised but not yet shipped, and the choice of what to open next. A promise to release is worth less than a release, and worth least in the weeks when ownership may change.
An older precedent, from December 2025, bears on this. On 15 December 2025 Nvidia announced it was acquiring SchedMD, the company that develops the Slurm job scheduler, which the post describes as used in more than half of the top 10 and top 100 systems on the TOP500 list. Nvidia pledged to keep developing and distributing Slurm as “open-source, vendor-neutral software.” Slurm’s license file says all its code and documentation is available under the GNU General Public License, with some contributed tools carrying other licenses and an OpenSSL linking exception. I found no published contract terms behind the pledge, and I am not suggesting it will be broken. The point is the structure. For Slurm, the safeguard is that anyone can fork the code. For weights, forking means fine-tuning a released checkpoint, which is real but is not the same as being able to build the next one. The pledge and the license do different jobs, and only one of them is enforceable by a stranger.
What I would check when the files appear
Reflection has set itself a deadline, so the grading can be mechanical.
When the weights ship:
- Is there a file on Hugging Face that a stranger can download without a waitlist, and does its license field and license text say Apache 2.0, unmodified?
- Do the technical report and model card arrive, and does the report say anything about the training data that meets the Initiative’s “sufficiently detailed information” bar, or only what the announcement already says, that the tokens are “diverse” and partly “curated”?
- Does the promised stack include training code, or only code to run, evaluate and fine-tune?
- Does anyone outside Reflection reproduce the benchmark table on the shipped checkpoint? The blog’s own competitors’ numbers come from third parties; its own should too.
- What has changed in the Nvidia story by then? That one can change in a single sentence, in either direction.
The valuation, which Laskin told CNBC in April was being raised at $25 billion pre-money according to the write-up, and the $800 million Nvidia has reportedly already committed, are both figures I could trace only to FT-derived coverage. In the coverage I read, neither company confirmed either figure, and I found nothing confirming that the round closed. I would not build on them.
This is a snapshot taken on 10 October 2026. By the end of the month either the files will exist or the word “later” will have stretched. In both cases the useful question is the same one: not whether the model is called open, but what a stranger can do with it on a Tuesday with no one’s permission.
References
- Reflection AI (2026). Introducing Beam. Company blog, 5 October 2026, updated 8 October 2026.
- Mugo, S. (2026). Nvidia in talks to acquire Reflection AI or increase investment, FT reports. Investing.com, 10 October 2026, citing the Financial Times.
- Reflection AI. Hugging Face organization page. Viewed the evening of 10 October 2026: no public models or datasets.
- Open Source Initiative. The Open Source AI Definition, version 1.0. Page shows no separate date.
- TechCrunch (2025). Reflection raises $2B to be America’s open frontier AI lab, challenging DeepSeek. 9 October 2025 (background).
- Nvidia (2025). NVIDIA Acquires SchedMD. Company blog, 15 December 2025 (background).
- SchedMD. Slurm license file (COPYING). GitHub, viewed 10 October 2026.
- Earlier essay: Mistral Large 4’s cyber lead comes from the test its rivals declined.