If Reflection AI's Beam is on your shortlist as the US-origin open-weight alternative to Kimi K3, you hold nothing yet that a change of owner can't touch. Reflection has promised Beam's weights under Apache 2.0 "later this month", and Nvidia is in early talks to buy the company, invest more in it, or hire its staff and license its technology, with an agreement possible within weeks. Once Apache 2.0 weights are published, that grant can't be withdrawn. Until they are, the licence is a sentence on a blog post, and the company that wrote it may soon answer to a chip vendor.
Don't standardise on Beam until a downloadable, Apache-licensed checkpoint exists in your own registry, and write your hosted-API terms around the deal structure Nvidia actually picks.
What Nvidia Is Reportedly Negotiating
Nvidia is weighing three structures for Reflection, according to the Financial Times as relayed by SiliconANGLE on October 11: an outright purchase, a larger stake possibly paired with more chips and computing power, or an acquihire in which Nvidia hires Reflection's staff and licenses its technology without taking control. People familiar with the talks told the FT a deal "could come within weeks" and that "the talks could still collapse." Neither company has commented.
Nvidia is already Reflection's anchor investor. SiliconANGLE reports that Nvidia led Reflection's $2 billion round in October 2025 and has put $800 million into the company in total, and that chief executive Misha Laskin told CNBC in April that the latest round closed at a $25 billion pre-money valuation. That October 2025 round was priced at $8 billion, with Sequoia, Lightspeed, DST and GIC among the other backers.
Reflection also carries a large compute bill. It agreed in June to pay SpaceX $150 million a month from July 1, 2026 through 2029 for Nvidia GB300 capacity at Colossus 2 near Memphis, a contract worth up to $6.3 billion that either side can end on 90 days' notice after the first three months. Nebius agreed in July to sell it another $1 billion of AI capacity. The FT's stake option, which pairs money with more chips, would speak directly to those commitments.
Beam's Weights Are Still a Promise
As of this week, Beam is a waitlisted API. You can't download it yet. Reflection's October 5 launch post describes a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, says it is in "final red-teaming," and offers access to "a select group of users" through a waitlist. The same post says "we will release the weights under an Apache 2.0 license," with the technical report and model card to follow later in October.
An independent check found the gap still open three days later. On October 8, eesel reported that Reflection's Hugging Face organisation listed zero models, that the OpenAI-compatible API (model ID Beam-501B-A23B) issues keys only once waitlist access is enabled, and that pricing sits behind a login. A search of the Hugging Face model API for Beam under a Reflection account on October 12 returned nothing.
That timing matters more than the benchmarks. Our October 6 comparison of Beam against Kimi K3 treated the Apache 2.0 licence as Beam's strongest procurement argument over the Chinese alternatives. That argument holds only for weights that have actually been published. A promise to publish binds nobody once the company making it has a new owner, a new board, or no research team.
Three Deal Shapes, Three Different Beams
Each structure on the table leaves you with a different counterparty, and Nvidia has already used all three in the past year.
| Structure | Nvidia precedent | Who you contract with afterwards | What changes for a Beam buyer |
|---|---|---|---|
| Acquihire + licence | Groq, December 2025 | Reflection, minus its senior team | The company stays, the people who build the next model may leave |
| Bigger stake + compute | Reflection itself, 2025 | Reflection, more dependent on Nvidia | Roadmap and supply tilt further toward one vendor |
| Outright purchase | Hugging Face, September 2026 | Nvidia, after regulatory review | Licence for future versions becomes Nvidia's call |
The acquihire would look like Groq. On December 24, 2025, Groq announced a non-exclusive licence of its inference technology to Nvidia. Founder Jonathan Ross, president Sunny Madra and other staff joined Nvidia, Groq stayed independent under a new chief executive, Simon Edwards, and the company said "GroqCloud will continue to operate without interruption." Neither company disclosed the price; CNBC reported $20 billion, against a prior Groq valuation near $7 billion. Nvidia ran the same play with Poolside in August, and as we noted then, no customer contract clause fired because ownership never moved. The FT says this shape would likely avoid a lengthy regulatory review, which makes it the fastest route for Nvidia and the hardest one for a customer to see coming.
A purchase would look like Hugging Face. Nvidia signed a definitive agreement on September 2, 2026 to buy Hugging Face for about $11.9 billion plus up to $1.0 billion in retention equity, with closing expected in the first half of 2027 subject to regulatory approvals. A Reflection purchase on similar terms would leave months between signing and close in which Reflection still operates on its own, and still controls whether the Beam release ships as announced.
A bigger stake changes the least on paper. Reflection keeps its team and its contracts. What shifts is leverage: a lab that rents Nvidia chips from SpaceX and Nebius, counts Nvidia as its lead investor, and then takes more of both has fewer reasons to disagree with Nvidia about anything.
What Apache 2.0 Protects, and What It Leaves Open
Apache 2.0 protects every copy of Beam that is published under it, permanently. Section 2 of the licence grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license", and Section 3 gives a matching patent grant that ends only for a licensee who sues claiming the work infringes a patent. No buyer of Reflection can claw back a checkpoint you have already downloaded under those terms.
It protects nothing else. A future owner can ship Beam 2 under a different licence, end the hosted API, change its price, or stop publishing weights altogether. Reflection says "Beam is the first model in a series" and that it is "already training what comes next". None of that work is covered by a licence on Beam 1.
Nvidia's own open models show what a different licence looks like. The NVIDIA Open Model License is commercially usable and allows derivative models, but Nvidia "may update this Agreement to comply with legal and regulatory requirements at any time," and your rights terminate automatically if you bypass a model guardrail without putting a "substantially similar Guardrail" in its place. That is a workable licence for most enterprises. It is also a licence Nvidia can amend, which Apache 2.0 can't be.
The Case That Nothing Changes
The strongest argument against worrying is that Nvidia wants open models to exist. It sells GPUs, and Beam was trained on Nvidia hardware: 6,144 GB300 GPUs for pretraining and 10.5K for reinforcement learning. A US-origin open model competing with DeepSeek and Kimi helps Nvidia's business whoever owns it. Reflection also built its commercial model on enterprise and sovereign-government customers; Laskin said in 2025 that the company would release weights while keeping datasets and training pipelines proprietary, and that research use would be free while revenue came from large enterprises and governments.
The argument is reasonable and may turn out right, but an incentive binds nobody, so your risk register should carry it as an assumption. If the weights ship under Apache 2.0 before any deal closes, most of the risk in this piece goes away for Beam 1. If they slip, every assumption in your shortlist rests on a release decision you don't control.
What to Do Before the Deal Lands
Get a published copy first, sign hosted terms second, and standardise only after you have both.
This Week:
- Put a hold on any architecture decision that names Beam as the default open-weight model. The shortlist can stay as it is.
- Ask your Reflection contact, in writing, for the date the Apache 2.0 weights will publish and where. Note whether the answer has changed since the October 5 post.
- Confirm that your fallback (Kimi K3, GLM, Mistral or Nemotron) still passes your provenance rules, so a slipped release does not stall the project.
This Month:
- When the weights appear, mirror them into your own registry the same day, with the licence file, model card and a checksum. We laid out the mirroring playbook when Hugging Face itself went up for sale.
- Size the hardware before anyone promises a self-hosted date. All 501 billion parameters must fit in memory even though only 23 billion are active; the same total-versus-active trap caught teams on Tencent's MoE.
- Run your own evals on the downloaded checkpoint. Treat Reflection's launch numbers as vendor claims; as of October 8, Beam was on neither Artificial Analysis nor Sierra's tau-bench leaderboard.
Before You Sign Hosted Beam Terms:
- Define change of control to include a licence-and-hire of key staff, because an ownership-only test would have missed both Groq and Poolside. The standard exit clauses we reviewed give between zero and 90 days.
- Ask for a deprecation notice period on the specific model version you build against, and a price hold through your first renewal.
- Get the right to run the same checkpoint yourself if the hosted endpoint changes owner or terms. With Apache weights in your registry, that right costs Reflection nothing to grant.
The Bottom Line
Nvidia took Groq's technology and leaders in December, Poolside's engineers in August and agreed to buy Hugging Face in September, each with a different deal shape. Buyers who wrote their protections around ownership alone missed the licence-and-hire version entirely. Beam adds a timing problem the earlier deals did not have: the asset buyers value most, an Apache 2.0 checkpoint of a 501-billion-parameter US-origin model, does not exist yet, and its release could come before the deal or after it.
If the weights publish first, mirror them and most of this risk is gone for Beam 1. If the deal lands first, read the licence on whatever ships before your team writes a line of code against it.
Continue Reading
- Reflection's Beam Beats Kimi K3 on One of 14 Shared Benchmarks
- Nvidia Hired 109 Poolside Engineers. No Clause Fired.
- Nvidia Buys the Bridge to Trainium. Go Grep Your Imports.
- Groq Runs Nvidia Now. Recount Your Non-Nvidia Capacity.
- Mistral Large 4 Beats Kimi K3 on Legal Work With No Licence Yet
- Best Air-Gapped LLM Stack: Apache Weights on vLLM Beat NVIDIA's Fee