American Open-Weight AI Models: Which US Labs Actually Ship Them
A landscape read on the US open-weight cohort — who these labs are, how they are funded, how far along each one is, and which buyer each one actually fits.
American open-weight AI models are now a real shortlist rather than a talking point. Six US developers publish, or have promised, models you can download and run yourself: Arcee AI, Poolside, Nvidia, Thinking Machines Lab, Meta, and Reflection AI. Two of those six shipped their first open weights only in 2026, and one has shipped nothing at all.
That matters because US buyers spent the last two years reaching for Chinese open-weight models on cost. DeepSeek, Qwen, and Kimi were simply the strongest weights you could download for the money. A US-jurisdiction option with a permissive license is a newer thing, and it is worth knowing which of these labs is a live option today versus a name to watch.
This page is the landscape read, not a spec sheet. It covers who each lab is, how it is funded, how mature its release is, and the buyer it fits. For a row-by-row safety and licensing table across every model on the market, including these US ones, use our AI model origins database — this page explains the labs behind those rows.
Which US labs ship open-weight AI models?
Six American developers publish open-weight AI models or have publicly committed to it: Arcee AI, Poolside, Nvidia, Thinking Machines Lab, Meta, and Reflection AI. Only five of them have weights you can download right now.
The cohort splits into two very different groups. Nvidia and Meta are large public companies that have shipped open weights for years and treat them as ecosystem strategy. Arcee AI, Poolside, Thinking Machines Lab, and Reflection AI are venture-backed startups founded since 2023 that treat open weights as their whole positioning.
The startups are the reason this page exists. Before 2026, a US buyer who wanted downloadable weights had two serious American answers: Llama and Nemotron. Now there are more, and they are licensed more permissively than Llama is.
Weighing an American open-weight model like Arcee Trinity Large or Nemotron against a cheaper Chinese one? We map the license, jurisdiction, and hosting cost to your actual compliance requirement before you commit hardware.
Book a ConsultationAmerican open-weight AI labs compared
The table below compares the six US labs on the five things a buyer actually decides on: license, whether you can self-host, how the lab is funded, how mature the release is, and who it fits. License and self-hostability come from each vendor's own published pages.
| Lab / model | Weights + license | Self-host? | Funding posture | Maturity | Best fit for |
|---|---|---|---|---|---|
| Arcee AI — Trinity (Large / Mini) | Open, OpenMDW 1.1 (launched under Apache 2.0), on Hugging Face | Yes | Venture-backed startup | Shipping since 2026, with a reasoning variant added later in the year | Regulated US firms that want a frontier-scale model with almost no license friction |
| Poolside — Laguna (S 2.1 / XS 2.1 / M.1) | Open, permissive but varies by model: OpenMDW 1.1 on S 2.1 and XS 2.1, Apache 2.0 on the earlier M.1 | Yes | Venture-backed startup, heavily capitalized | Shipping since July 2026 | Engineering teams that want private, local agentic coding |
| Nvidia — Nemotron 3 | Open weights, plus published training data; NVIDIA Open Model License or OpenMDW-1.1 by variant | Yes | Public company; funds the ecosystem rather than raising from it | Mature; multiple generations shipped | Teams building agents on Nvidia hardware who want a size range from 30B to 550B |
| Thinking Machines Lab — Inkling | Open, Apache 2.0, on Hugging Face | Yes, with serious hardware | Venture-backed startup founded by former OpenAI CTO Mira Murati | First open model shipped July 2026 | Teams whose plan is to fine-tune a broad base model rather than use it as-is |
| Meta — Llama | Open weights under Meta's own community license | Yes | Public company | Mature, with the largest tooling ecosystem | Teams that want maximum ecosystem support and can live with community-license terms |
| Reflection AI | None released | No | Company-confirmed $2B raise; Nvidia among its investors | Pre-release | No one yet — watch it, do not plan on it |
Read the license column before the flag column. Apache 2.0 and OpenMDW 1.1 are both permissive, so commercial deployment is low friction under either — but they are different documents, and an approval for one does not automatically carry to the other. Meta's community license is more open than a closed API but carries usage and naming terms you should read at scale.
Arcee AI: the smallest lab with the biggest open release
Arcee AI is the US lab furthest along on general-purpose open weights. It released Trinity Large with the weights published on Hugging Face, making it one of the largest general-purpose open models released by a US company. Check the license on the model card before you plan around it: Trinity launched under Apache 2.0, and Arcee later moved the whole family to OpenMDW 1.1, the Linux Foundation's model license, applied retroactively. Both are permissive, but they are not the same document.
The story behind it is the interesting part. Arcee was founded in 2023 by Mark McQuade, Jacob Solawetz, and Brian Benedict, and remains a small, venture-backed lab. Arcee reports training the Trinity lineup on 2,048 Nvidia Blackwell B300 GPUs for around $20 million. It does not publish a headcount or a funding total on its own site, so treat third-party figures as estimates.
That budget is the point. It is a rounding error next to what the largest labs spend on a frontier run. Arcee proved a small US team can put frontier-scale open weights on the table, which is why every other name on this page now has a harder story to tell about needing billions first.
Who Arcee fits: a US or regulated buyer who wants a large, capable model under a license with almost no strings, hosted on infrastructure they control. Who it does not fit: a team that needs a deep tooling ecosystem, vendor support contracts, or a long track record. Arcee is new, and its ecosystem is thin next to Llama's.
Start with our Trinity Large explainer for specs and access paths, and the Trinity Large alternatives page if you are shortlisting against other open-weight options.
Poolside: open weights for coding, with a license that varies by model
Poolside is a US-headquartered lab building agentic coding models, and it releases some of them with open weights. Its Laguna family shipped publicly in July 2026. Poolside's own models page states that it releases the family under permissive licenses ranging from OpenMDW to Apache 2.0, and adds that licenses vary by model.
That last phrase is the one to act on. The Hugging Face model cards show OpenMDW 1.1 on both Laguna S 2.1 and Laguna XS 2.1, and Apache 2.0 on the earlier M.1. Read the card for the exact variant you intend to deploy, because the answer is not the same across the family.
Poolside's positioning is narrow on purpose. These are coding models, sized so the small variant runs locally on a single machine. That is a genuinely useful shape for firms that will not send source code to a hosted API.
Who Poolside fits: engineering teams that want private, on-machine coding assistance and are comparing it against a hosted IDE assistant. Who it does not fit: anyone looking for a general-purpose business model — Poolside is not competing for that job. Our Poolside Laguna explainer covers the family in detail.
Nvidia and Thinking Machines Lab: the two already shipping at scale
Nvidia and Thinking Machines Lab are the two US labs shipping open weights at real scale today, for very different reasons. Nvidia publishes the Nemotron family with open weights and, unusually, open training data. Thinking Machines Lab published Inkling, its first open-weight model, in July 2026 under Apache 2.0.
Nemotron 3 is the most practical US open-weight family for most buyers, because it comes in sizes. The lineup spans a compact 30B-class model up to a 550B-class one, so you can match the model to the hardware you actually have. Licensing runs under the NVIDIA Open Model License or OpenMDW 1.1 depending on the variant, both permissive.
Nvidia's motive is not subtle. Every open model that runs well on its hardware sells more of its hardware. That alignment makes Nemotron unusually stable as a long-term choice — Nvidia has no commercial reason to close it.
Inkling is the opposite bet. It is a very large mixture-of-experts model with a small active slice, designed to be customized rather than to top a leaderboard. The lab says so plainly, and pairs it with its own fine-tuning platform.
Who each fits: Nemotron fits teams that want a permissively licensed model in whatever size their GPUs can actually hold. Inkling fits teams whose plan is fine-tuning on proprietary data, not using a base model as-is. We cover both in depth in the Nemotron 3 explainer and the Thinking Machines Inkling business guide.
Reflection AI: enormous funding, nothing to download
Reflection AI has raised more money than any other name on this page and has released no open weights. The company was founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou. Its own site states the mission plainly: make intelligence open and accessible to all.
The funding is unusually large for a pre-release lab. Reflection has confirmed a $2 billion raise on its own site and lists Nvidia among its investors. Press through 2026 has described further raises, valuations, and a large compute agreement — we are not restating those figures here, because they are reported rather than confirmed on Reflection's own pages, and they move.
What matters for a buyer is simpler. There is no model card, no weights file, and no published benchmark to evaluate. Nothing about Reflection AI can be piloted today.
Who Reflection fits: no one, yet. Put it on a watch list, set a calendar reminder to check its site each quarter, and make no roadmap commitments until weights and a license exist. Our Reflection AI explainer tracks what is verified and what is still unannounced.
This is the honest asymmetry in the American open-weight story. The best-funded entrant has shipped nothing, and the smallest one has shipped the most.
Meta and Llama: still the default, no longer the safe long-term bet
Llama is still the most widely deployed American open-weight family, and it is still the easiest to hire for. Nearly every serving stack, quantization tool, and fine-tuning framework supports it first. For a team that values tooling maturity over license purity, that is a real advantage.
Two caveats belong on the record. First, Llama ships under Meta's own community license, not Apache 2.0 or MIT. It permits commercial use but carries acceptable-use terms, a naming requirement for derivatives, and a threshold that only affects very large platforms.
Second, Meta's long-term commitment to open weights has been questioned throughout 2026 in press coverage of its superintelligence group and its closed releases. We are not treating those reports as settled fact. But a buyer planning three years out should check llama.com directly rather than assume the open line continues at frontier scale.
Who Llama fits: teams that want the widest ecosystem and can accept community-license terms. Who should look elsewhere: anyone whose legal team wants a standard permissive license, or whose procurement process treats license ambiguity as a blocker. Arcee's OpenMDW-licensed weights and Nemotron's permissive licenses both clear that bar more cleanly.
Are American open-weight models better than the Chinese ones?
No — not automatically, and not on capability per dollar at the top end. The strongest open-weight models a business can download in 2026 still include Chinese families. Kimi K3 from Moonshot AI, DeepSeek, and Qwen have led on published open-weight results and on cost for long enough that this is not a close call to pretend otherwise.
What changed is not the leaderboard. It is that a US buyer with a jurisdiction requirement now has a permissively licensed American option that is frontier-scale rather than a compromise. Arcee Trinity Large and Nemotron 3 both clear that bar. A year ago the honest answer was Llama or nothing.
So the question splits in two. If your requirement is the most capable open weights you can self-host for the lowest cost, the Chinese families remain in the running and often win. If your requirement includes a US-jurisdiction developer, permissive licensing, and a supply chain your board will sign off on, the American cohort is now a real shortlist.
One point deflates a common argument on both sides. If you download the weights and run them on your own hardware, the developer's country stops controlling your data — the model has no phone home. Jurisdiction risk on an open-weight model is mostly about the hosted API path, not the self-hosted one. Our Trinity Large vs DeepSeek comparison works through that head-to-head, and US alternatives to Chinese AI models covers the swap decision more broadly.
Why Nvidia sits behind almost all of this
Nvidia is the connective tissue of the American open-weight cohort, as a chip supplier, an investor, and a lobbyist. It ships its own Nemotron open models. It led Reflection AI's reported $2 billion round. And every model on this page was trained on its GPUs.
It has also taken a public policy position. In July 2026, Nvidia published an open letter titled "Open Weights and American AI Leadership," arguing against premature restrictions on open models that would stifle competition or push development overseas. Roughly two dozen companies signed on within days.
The commercial logic is worth stating plainly rather than treating the letter as neutral. Open models running on Nvidia hardware sell Nvidia hardware. That does not make the argument wrong, but a buyer should read it as an interested party's position, not an impartial safety assessment.
The practical takeaway: the American open-weight lane is real, but it is currently underwritten by one company's strategy. That is a concentration risk worth naming in a vendor review, even though it also means the lane is unlikely to disappear soon.
What a buyer should actually do now
Pick your requirement before you pick your model, because the requirement decides the shortlist. Most teams evaluating American open-weight AI models are answering one of three different questions, and each has a different right answer.
In the model-selection work we do for regulated firms, "American" is almost never the actual requirement written down anywhere. What is written down is a contract clause about where data may be processed, or a board question about who could be compelled to hand it over. Both are answered by self-hosting under a permissive license, not by the country on the model card — and we see the same thing running our own automation fleet across a portfolio of sites, where a model swap comes down to license friction and cost per run.
That is why the license column above does more work than the flag column. Get it wrong and the problem surfaces in a procurement review, months after the deployment is built.
- If the requirement is "permissive license, US developer, capable today" — evaluate Arcee Trinity Large and Nvidia Nemotron 3 first. Both ship permissive licenses and downloadable weights.
- If the requirement is "private coding assistance that never leaves our machines" — evaluate Poolside Laguna's open variants, and confirm the license on the exact variant you plan to run.
- If the requirement is "fine-tune a base model on our own data" — Inkling was built for that job, and Arcee's OpenMDW terms permit it without restriction.
- If the requirement is "maximum ecosystem and tooling" — Llama is still the answer, provided the community license clears your legal review.
- If the requirement is "the strongest open weights per dollar, full stop" — keep the Chinese families in the evaluation and compare honestly. See what open-weights models are and the real cost of open-weights models before you size hardware.
- In every case, verify the license on the vendor's own page for the exact variant. Family-level license claims are frequently wrong, including in press coverage.
Conclusion: the American open-weight shortlist is real, and short
American open-weight AI models went from one credible option to several in a single year. Arcee AI ships frontier-scale permissively licensed weights from a small lab. Nvidia ships Nemotron across a wide size range. Thinking Machines Lab shipped Inkling for teams that will fine-tune. Poolside covers private coding. Llama remains the ecosystem default, and Reflection AI remains a name with funding and no product.
The cohort is real but early, and it has not overturned the capability-per-dollar case for the leading Chinese open-weight families. What it has done is give a US buyer with a jurisdiction or licensing constraint a genuine choice instead of a compromise.
Choose on license, size, and what your hardware can actually hold. Verify every license claim on the vendor's page. Then pilot on your own workload, because that is the only test that predicts what you will live with.
If you want help running that evaluation — or standing up a self-hosted open-weight deployment that survives a compliance review — that is the work Layer3 Labs does for small and mid-size firms in regulated industries.
Frequently Asked Questions
- The US labs shipping downloadable open weights in 2026 are Arcee AI (Trinity Large and Trinity Mini), Poolside (the open Laguna variants), Nvidia (the Nemotron family), Thinking Machines Lab (Inkling), and Meta (Llama). Reflection AI is funded to compete in this space and states an open-model mission, but has not released weights. Licenses vary by lab and by model variant, so verify the exact variant before deploying.
- Thinking Machines Lab's Inkling ships under Apache 2.0, and Arcee's Trinity models ship under OpenMDW 1.1 after launching on Apache 2.0 — both are about as low-friction as commercial licensing gets. Nvidia's Nemotron variants use either the NVIDIA Open Model License or OpenMDW 1.1, both permissive. Meta's Llama is the outlier among the established US options, shipping under a vendor community license with usage and naming terms.
- Not on raw capability per dollar at the top end, where the leading Chinese open-weight families have held an edge through 2026. The American cohort's advantage is jurisdiction and licensing clarity, not benchmark superiority. If your only requirement is the strongest self-hostable weights for the lowest cost, evaluate both sides honestly rather than assuming the US option wins.
- No. As of August 2026 Reflection AI has published no open weights, model card, or license, despite reported funding of more than $2 billion including a round led by Nvidia. Its own site states a mission to build open models but announces no release date. Treat it as a company to monitor, not a model to evaluate.
- Yes, and self-hosting is what actually solves the data-residency problem. When you download weights and run them on hardware you control, the model does not send data anywhere — the developer's country stops mattering for your inputs. Jurisdiction risk on an open-weight model mostly attaches to the vendor's hosted API path, not the self-hosted one.
- Nvidia supplies the GPUs every one of these models was trained on, invests in several of the labs, and publishes its own Nemotron open models. It led a reported $2 billion round in Reflection AI and, in July 2026, published an open letter arguing against premature restrictions on open models. The commercial logic is direct: open models running on Nvidia hardware sell Nvidia hardware, so read its policy position as an interested one.
Not sure which American open-weight model fits your requirement?
Layer3 Labs helps SMBs and regulated firms turn a vague "we need a US model" into a specific license, jurisdiction, and hosting decision — then stands up the self-hosted deployment on infrastructure they control.
Book a free open-weights assessment