Open-Weight AI in 2035: Scenarios and Falsifiers

Two axes, not four separate bets

Ask four people building on open-weight models where the ecosystem will be in 2035 and the answers cluster into four seemingly independent bets. Will open-weight models close the capability gap with closed frontier systems, the way the gap Epoch AI tracks has already narrowed from roughly a year to an average of about three months ^[1] — or does a persistent gap remain, the kind the UK’s AI Security Institute still measures at four to seven months on the tasks it worries about most ^[2]? Will regulatory regimes start treating an open-weight release as needing the same safety review as a closed deployment, the way the EU AI Act already strips its open-source exemption from any model whose training compute exceeds ten-to-the-twenty-fifth floating-point operations ^[5] — or does most regulation stay light-touch, the way California’s frontier-AI law applies by compute alone with no reference to how a model is distributed ^[6], and the way the White House’s own AI Action Plan promotes open release for its “geostrategic value” rather than subjecting it to extra scrutiny ^{[7, 8]}? Will the ecosystem consolidate around a few dominant families, the way adoption already concentrates on a short list of providers even as new entrants keep arriving ^[17] — or does it keep fragmenting across many providers, the way a single mid-2026 snapshot already counts Meta, Mistral, DeepSeek, Alibaba’s Qwen, Moonshot, Zhipu, MiniMax, Cohere, Reflection AI and Thinking Machines Lab all shipping competitive open weights within months of one another ^[3]? And will open-weight licensing converge on a genuinely open-source-compatible standard, the way Qwen, Mistral, gpt-oss, DeepSeek’s newest weights and Zhipu’s GLM line have all now settled on plain Apache 2.0 or MIT text ^{[13, 14, 10, 11, 18]} — or does custom, restrictive licensing keep proliferating, the way Meta’s Llama and Moonshot’s Kimi both still layer a bespoke commercial-scale clause on top of, or instead of, standard licence text ^{[9, 15]}?

Argued one at a time, these read as four independent coin flips. They are not, quite. The second and fourth bets are the same underlying question approached from two directions — whether external standardization arrives through public law or through private contract — so this article treats them as one axis. The third bet, about market structure, is not independent of the other two at all; it is downstream of them, and the section after next argues that explicitly rather than assuming it. That leaves two axes. Axis A, capability, asks whether the gap between the strongest open-weight models and the closed frontier keeps narrowing toward rough parity on most tasks, or persists at a stable distance. Axis B, governance, asks whether regulatory regimes and licensing norms both converge on treating open-weight release as needing the same kind of external scrutiny closed deployment already gets, or whether both stay light-touch and fragmented — law by narrow, easily-cleared carve-outs, licences by bespoke bilateral terms written one release at a time. Crossing them gives four scenarios, each with a horizon, assumptions, observable indicators, and an explicit disconfirmation condition. None is named as the likely outcome; ecosystem consolidation is folded back in once its dependence on the other two is established.

The documented present

Fact. The best-characterised measure of the capability gap comes from Epoch AI, which defines a composite Epoch Capabilities Index across many benchmarks and calculates the horizontal time-distance between an open-weight frontier and a closed one. As of its most recent published analysis, frontier open-weight models lag the closed state of the art by an average of roughly three months, with a ninety-percent confidence interval of about one to five months, and a vertical gap of around seven ECI points — comparable to the distance between two recent generations of a single closed model family. A year earlier the same measure put the lag closer to a full year ^[1].

Fact. Epoch’s own account is careful to flag two things that cut against a simple “gap closing” headline. Public benchmarks may flatter open models because their test sets are easier to optimise against once released; and the ECI methodology captures weight performance specifically, leaving aside the separate, less encouraging question of whether an adopter can also match a closed lab’s data, infrastructure and training signal advantages just by downloading a checkpoint ^[1].

Fact. A domain-specific measurement tells a related but not identical story. The UK’s AI Security Institute ran seventy narrow cyber tasks and a set of autonomous attack-range exercises against the strongest available open-weight models and reported, in a report published in July 2026, that the leading open models trailed the closed cyber-capability frontier by four to seven months — narrower than the six-to-ten-month gap the same institute measured through most of 2025. On specific narrow tasks, one open model performed comparably to a closed model released four to five months earlier; on the harder cyber-range exercises, the gap widened to about seven months. The same report noted a large cost asymmetry: one open model completed tasks at roughly twenty-eight US cents each against roughly twelve US dollars and fifty cents for a comparable closed model ^[2].

Fact. A third, independent snapshot comes from Mozilla’s inaugural State of Open Source AI report, published in July 2026, which compared the strongest closed and open models on its own intelligence measure and found the closed model scored 61 against 57 for the strongest open model — a difference the report characterises, citing Epoch’s own capability index, as “about one release cycle” ^[3]. Artificial Analysis’s own live leaderboard shows a similar picture from a different angle: as of this writing its top-ranked open-weight model scores 60 on its Intelligence Index, with several others close behind in the low-to-mid fifties, and its own published chart of “Progress in Open Weights vs. Proprietary Intelligence” shows the trend narrowing over successive quarters rather than holding steady ^[4].

Fact. These three measurements agree on direction and disagree on magnitude — three months on Epoch’s general index, four to seven months on AISI’s cyber-specific one, “about one release cycle” on Mozilla’s — and that disagreement is itself informative. There is no single, agreed unit for “how far behind” an open-weight model is; the honest summary is that the gap is narrow enough to be measured in months rather than years, narrowing further on most read­ings, and not yet reliably zero on any of them.

Fact. On regulation, the EU AI Act’s Article 53 exempts providers of models released under a free and open-source licence, with publicly available weights, architecture information and usage information, from the documentation and downstream-transparency obligations that otherwise apply — but states plainly that “this exception shall not apply to general-purpose AI models with systemic risks” ^[5]. Systemic-risk status attaches, under a companion provision, once a model’s training compute exceeds ten-to-the-twenty-fifth floating-point operations — a threshold that a growing number of the largest open-weight releases now sit close to or above, which means the exemption is narrower in practice than its headline description suggests.

Fact. California’s Transparency in Frontier Artificial Intelligence Act, signed in September 2025 and effective from January 2026, takes a different design. It defines a “frontier model” purely by training compute — more than ten-to-the-twenty-sixth floating-point operations, cumulative across the original run and any subsequent fine-tuning — and its disclosure, incident-reporting and risk-framework obligations apply to any developer crossing that threshold, with no separate carve-out or discount for models whose weights are subsequently published ^[6]. Where the EU’s rule gives openness a conditional exemption, California’s gives it none: the two jurisdictions are not converging on a single answer to whether open-weight release warrants lighter treatment.

Fact. Federal US policy currently pulls in the opposite direction from either regulatory design. The White House’s July 2025 “Winning the Race: America’s AI Action Plan” devotes a section to encouraging open and open-weight models, and states — as a policy assertion, not a settled empirical claim — that “open source and open-weight models could become global standards in some areas of business and in academic research worldwide. For that reason, they also have geostrategic value,” directing agencies including the National Telecommunications and Information Administration to convene stakeholders specifically to drive adoption of open models by smaller businesses ^{[7, 8]}. That is a strategic argument for less friction on open release, not more, sitting in the same policy landscape as California’s compute-only threshold and the EU’s conditional exemption.

Fact. On licensing, the highest-profile 2025 releases moved toward, not away from, standard text. OpenAI’s gpt-oss-120b and gpt-oss-20b shipped under a plain Apache License 2.0 ^[10]. Alibaba’s Qwen3 line ships under Apache 2.0 as listed on its own model card ^[13]. Mistral AI’s December 2025 Mistral 3 family — from the three-billion-parameter Ministral up to the six-hundred-seventy-five-billion-parameter Mistral Large 3 — states plainly that “all models are released under the Apache 2.0 license” ^[14]. Zhipu AI’s GLM-4.5 line lists a plain MIT licence on its own model card ^[18]. DeepSeek’s newest weights, including the V3.2 line, ship under a standard MIT licence ^[11] — a genuine change from the company’s own October 2023 DeepSeek License Agreement, Version 1.0, which carried behavioural-use restrictions on military use, automated decisions affecting legal rights, and generation of disinformation, and required derivatives to carry the same restrictions forward ^[12].

Fact. Two prominent 2025 families moved the other way, or did not move at all. Meta’s Llama 4 Community License Agreement, effective April 2025, keeps the clause that has defined the Llama licence since 2023: any licensee whose products or services exceed seven hundred million monthly active users “must request a license from Meta,” which Meta may grant in its sole discretion, and any model trained, fine-tuned or otherwise improved using Llama materials must include “Llama” at the start of its name if distributed ^[9]. Moonshot AI’s Kimi K2 shipped under a “modified MIT” licence that layers a bespoke clause onto the standard text: any commercial product or service built on it with more than one hundred million monthly active users, or more than twenty million US dollars in monthly revenue, must “prominently display ‘Kimi K2’” in its interface ^[15]. Mozilla’s own report notes that Moonshot’s licensing has not even settled from one release to the next — Kimi K2 used “modified-MIT,” and, in the report’s words, “that did not settle” for Kimi K3, which shipped under a different custom text again ^[3].

Fact. The definitional yardstick for whether any of this counts as genuinely open sits with the Open Source Initiative’s Open Source AI Definition, published in October 2024. It specifies four freedoms — to use, study, modify and share the system for any purpose without permission — and three required components for the “preferred form” of a compliant release: sufficiently detailed information about training data “so that a skilled person can build a substantially equivalent system,” the complete source code used to train and run the system, and the model parameters themselves ^[16]. None of the licence texts above, plain Apache 2.0 and MIT included, addresses that first component; a model can be released under an unimpeachably standard software licence while still falling short of the Definition’s own bar for open source AI, because the licence governs the weights, not the training data behind them.

Fact. On market structure, Nathan Lambert’s ongoing tracking of Hugging Face fine-tune and download activity found that, despite a marked increase in the number of organisations releasing open-weight models through 2025, “the share of finetuned models has concentrated” — with new entrants including Z.ai, MiniMax, Nvidia and Moonshot individually registering as, in his words, “a rounding error in adoption metrics” next to a small group of leaders. The same analysis found the geographic centre of gravity had moved decisively: “2025 saw the end of Llama and Qwen triumphantly took its spot” as the most-adopted open-weight family ^[17]. More organisations are releasing competitive weights than at any point before; adoption is not spreading evenly across them.

Why these two axes, and why ecosystem consolidation is a consequence, not a third

Axis A is consequential because its two poles decide what an “open-weight model” is even competing to be: a genuine substitute for a closed frontier system on most tasks, the way Mozilla’s 57-against-61 reading and Artificial Analysis’s narrowing chart both suggest is close ^{[3, 4]} — or a permanently trailing option whose value lies in cost and control rather than raw capability, the way AISI’s own cyber-specific gap and Epoch’s own caveats about data and infrastructure moats both suggest is still real ^{[2, 1]}. It is uncertain because three independently run measurements, published within months of one another in 2025 and 2026, do not agree on the size of the gap even as they agree on its direction.

Axis B is consequential because it decides whether “open-weight” names a release format that draws the same external scrutiny a closed deployment gets, or one that is treated, in law and in contract, as fundamentally lower-stakes. It is uncertain because the current record pulls in different directions inside the same eighteen months: the EU’s conditional exemption disappears entirely above a compute threshold ^[5], California’s threshold applies with no exemption at all ^[6], and the White House’s own policy pushes toward less friction rather than more ^[7]. Licensing shows the same split: five prominent 2025 families converged on plain Apache 2.0 or MIT text ^{[10, 13, 14, 18, 11]}, while two others — one Western, one Chinese — kept or introduced bespoke commercial-scale clauses, with one of them changing its own approach release to release ^{[9, 15, 3]}.

Whether the open-weight ecosystem consolidates around a handful of dominant families or stays fragmented across many providers is not a third axis; it is what the other two jointly produce. Write $O(t)$ for a stylised index of how close the strongest open-weight models sit to the closed frontier — Axis A’s own proxy, anchored in the Epoch and Artificial Analysis readings above ^{[1, 4]} — and $W(t)$ for the share of open-weight releases operating under standardized external governance: a regulatory regime that reviews them the way closed deployment is reviewed, or a licence that is genuinely OSI-compatible rather than bespoke ^[16]. Reaching and holding a position near the frontier takes capital; clearing a standardized compliance and licensing bar on top of that takes more. A family that is merely near-frontier but unregulated and loosely licensed can still be undercut by a cheaper, less scrutinised rival — which is closer to today’s fragmented reality, where adoption concentrates on a handful of leaders for reasons of quality and trust even though barriers to entry remain low ^[17]. A family that clears a standardized compliance bar but sits well behind the frontier has little to sell. Durable consolidation around a few dominant, well-capitalised families needs both conditions cleared at once, not an average of them:

$K(t)$ stays low however high either term climbs alone. Axis A determines whether $O(t)$ can plausibly clear $o^{*}$ within this article’s horizon; Axis B determines whether $W(t)$ can. Neither can be inferred from the other, which is why they are kept as two axes rather than folded into one — and why a narrowing capability gap sitting beside a licensing and regulatory picture that is still openly split, as documented above, does not yet settle which way the ecosystem’s structure moves.

Four scenarios toward 2035

Crossing the two axes gives four cells. None is named as the likely outcome; each is named for the instrument in this article’s own room that best captures its logic.

Scenario one: The Stamped Jaws — capability converges, governance converges

Mechanism. The narrowing trend Epoch AI, AISI and Mozilla each already document ^{[1, 2, 3]} continues until the strongest open-weight models sit within a small, stable margin of the closed frontier on most tasks, consistent with Artificial Analysis’s own narrowing chart ^[4]. In parallel, regulators generalise the EU’s compute-triggered systemic-risk test ^[5] into something that also weighs how capable a release actually is rather than compute alone, and licensing settles the way DeepSeek’s own move off its 2023 custom terms ^[12] onto plain MIT ^[11], alongside Qwen’s, Mistral’s and gpt-oss’s Apache 2.0 releases ^{[13, 14, 10]}, already suggests is achievable at scale.

Horizon. Recognisable capability convergence on general tasks by 2030–2031; a settled, externally verified governance layer plausible by 2033–2035.

Assumptions. The narrowing documented by Epoch and Artificial Analysis continues rather than stalling at today’s few-month, single-digit-point residual; a regulator becomes willing to write a rule referencing capability proximity rather than compute alone; a critical mass of releasers accept that a compliance cost is worth paying for the market access it buys.

Observable indicators. Third-party capability trackers report the open-closed gap at or near zero on general benchmarks for two consecutive years; a majority of the ten most-downloaded open-weight families ship under a plain OSI-approved licence with no bespoke commercial-scale clause; at least one jurisdiction extends binding safety-review obligations to open-weight releases specifically because of their capability, not only their compute.

Disconfirmation. Falsified if, by 2032, the capability gap Epoch and AISI track has stopped narrowing for two consecutive readings, or if no jurisdiction has moved past a pure compute threshold to a rule that also weighs how capable a release actually is.

Scenario two: The Free Jaws — capability converges, governance stays fragmented

Mechanism. Capability keeps converging as in scenario one, but governance does not follow it. The EU’s exemption keeps working as designed for everything under its compute threshold ^[5], California’s law keeps applying by compute alone without ever specifically targeting open releases ^[6], and federal US policy keeps actively promoting open release as a strategic asset ^{[7, 8]} rather than treating it as a risk requiring closer management. Licensing keeps splitting the way it already does: most releasers converge on Apache 2.0 or MIT because it is simpler to adopt something that already exists, while at least one or two prominent families keep writing bespoke clauses the way Meta and Moonshot already do, with at least one of them still revising its own approach release to release ^{[9, 15, 3]}.

Horizon. Capability convergence recognisable by 2029–2030; governance staying at roughly today’s fragmented baseline through 2035.

Assumptions. Regulators continue to find compute the only tractable, hard-to-game trigger, and no publicly attributed incident forces a rethink; the economic case for a standard licence — lower legal friction for adopters — continues to outweigh the strategic case for a bespoke one for most, though not all, releasers.

Observable indicators. Capability trackers show the gap near zero while no jurisdiction adds openness- or capability-contingent review beyond compute; the licence landscape stays a mix, with most top releases on standard text and a persistent minority on custom terms.

Disconfirmation. Falsified if governance converges at the thresholds scenarios one or three define — that would move the world out of this cell.

Scenario three: The Weighed Gap — capability gap persists, governance converges

Mechanism. The gap AISI already measures at four to seven months on the tasks it cares about most ^[2] stabilises rather than continuing to close — plausible if the hardest residual capability, the kind that shows up on the most demanding reasoning and highest-stakes tasks, keeps requiring resources only a few closed labs commit, in line with Epoch’s own caveat about data and infrastructure moats that weight performance alone does not capture ^[1]. Governance converges anyway: regulators extend compute-triggered obligations downward or add capability-linked triggers regardless of whether the gap keeps closing, and licensing settles onto standard terms the way it has already begun to across most of the field ^{[13, 14, 10, 11, 18]}.

Horizon. A stabilised, non-closing gap recognisable by 2029; a converged governance layer plausible by 2032–2034.

Assumptions. The residual capability gap reflects something structural — training-scale economics, proprietary data, or post-training expertise that open releases cannot replicate cheaply — rather than a temporary lag that further engineering closes; regulators decide that a stable, moderate gap is not, by itself, a reason to relax scrutiny.

Observable indicators. The gap Epoch and AISI track stops narrowing for two consecutive years while stabilising rather than widening; binding regulation nonetheless extends further into the open-weight release pipeline; licence text keeps consolidating onto standard terms even as the underlying models stay a measurable step behind.

Disconfirmation. Falsified if the capability gap resumes narrowing toward the thresholds scenario one defines — that would move the world toward scenario one instead.

Scenario four: The Scattered Bins — capability gap persists, governance stays fragmented

Mechanism. Neither pressure resolves. The gap stabilises or narrows only slowly, roughly the way it has moved over the past two years without yet reaching zero on any of the three measurements above ^{[1, 2, 3]}, and governance stays close to where it is today: a narrow, compute-gated exemption in the EU that most releases still qualify for ^[5], no open-weight-specific treatment in California ^[6], active federal promotion in the US ^[7], and a licensing landscape that keeps mixing standard and bespoke terms the way it already does ^{[9, 15, 13, 14]}. This scenario’s leading indicator is not a future event; it is the documented present above, continuing.

Horizon. Close to today’s baseline; recognisable as the stable case by 2028 if none of the trends above accelerate; could persist through 2035.

Assumptions. No publicly attributed incident forces emergency regulation of open releases specifically; no single technical breakthrough closes the residual capability gap faster than the historical pace.

Observable indicators. Published capability-gap figures stay in roughly today’s range; no jurisdiction adds capability-linked review beyond compute; the ten most-downloaded families keep splitting between standard and bespoke licence text roughly as they do today.

Disconfirmation. Falsified if either capability convergence or governance convergence is observed at the thresholds scenarios one through three define.

What all four share, and the wildcard neither axis names

Three things hold across every cell. First, a meaningfully sub-frontier tier of open-weight releases survives in all four scenarios, including the fully convergent one — smaller, cheaper, more customisable models remain worth releasing and adopting for cost, latency and control reasons even where a frontier-adjacent open model also exists, exactly as today’s field already contains both frontier-chasing releases like DeepSeek’s and Qwen’s largest models and much smaller, purpose-built ones. Second, which specific family leads the capability race — Qwen, DeepSeek, a resurgent Llama, or an entrant not yet founded — is a narrower, more replaceable detail than the axis itself; the axis is compatible with any winner, the way leadership has already rotated from Llama to Qwen once inside this article’s own research window ^[17]. Third, none of the four requires closed frontier capability to plateau; Axis A describes relative distance, not an absolute ceiling, so a closed frontier that keeps advancing is compatible with every scenario above as long as open-weight models advance at a comparable or faster pace.

The four scenarios share a blind spot too. All four assume the roughly gradual movement the 2025–2026 evidence already shows. A large-enough shock would not fit cleanly into any of them. A single incident in which an open-weight model’s removed safety guardrails were shown to have materially contributed to real-world harm — the exact scenario AISI’s own cyber-capability tracking exists to give early warning of ^[2] — could force emergency, capability-linked regulation well ahead of scenario three’s gradual path, and could simultaneously push licensing toward standardisation as an industry self-preservation response, moving both axes at once rather than along either scenario’s traced course. Equally, a sustained run of years with no such incident, combined with the capability gap continuing to narrow on Epoch’s and Artificial Analysis’s own readings ^{[1, 4]}, could remove whatever practical urgency currently exists for tighter regulation, holding Axis B at its fragmented pole even as Axis A converges — which is scenario two, arrived at not gradually but because the case for scenario one’s regulatory half never gets made.

Two predictions, stated separately from the scenarios

Prediction one. Horizon: end of 2029. A majority of newly enacted, binding frontier-AI statutes in G7 jurisdictions will follow California’s design — a compute threshold applying regardless of release openness ^[6] — rather than the EU’s design, which strips its exemption only above a systemic-risk threshold but otherwise treats open release as warranting less documentation ^[5]. Assumption: regulators default to the simpler, harder-to-game trigger once the practical difficulty of assessing “genuine openness” case by case becomes apparent from EU implementation experience. Indicator: the text of newly enacted frontier-AI statutes and their treatment, or non-treatment, of open-weight release. Disconfirmed if a majority of statutes enacted by G7 jurisdictions before the horizon include an EU Article 53-style open-source carve-out rather than a pure compute test.

Prediction two. Horizon: end of 2030. Among the ten most-downloaded open-weight model families, a majority will ship under a plain OSI-approved licence — Apache 2.0 or MIT — with no additional bespoke commercial-scale clause, continuing the shift already visible in DeepSeek’s move off its 2023 custom licence ^{[12, 11]} and in Qwen’s, Mistral’s and gpt-oss’s current releases ^{[13, 14, 10]} — while at least one top-ten family continues to layer a bespoke threshold clause onto an otherwise standard text, in the pattern Meta and Moonshot already show ^{[9, 15]}. Assumption: the legal-friction advantage of a licence adopters can evaluate once continues to outweigh, for most releasers, whatever commercial protection a bespoke clause offers. Indicator: the licence field on the ten most-downloaded model repositories tracked by Hugging Face download volume. Disconfirmed if a majority of the top ten instead ship under a bespoke, non-OSI-approved licence by the horizon.

What to take away

The refusal to pick a favourite among these four cells is the substantive claim, not a hedge around one. In August 2026 the evidence is genuinely split on both axes at once: an Epoch AI reading that puts the capability gap at about three months sits beside a UK government reading that puts a narrower slice of the same gap at four to seven months, both published within the same twelve months and both showing the same direction of travel without agreeing on its size ^{[1, 2]}. An EU rule that lets open release skip real obligations below a systemic-risk threshold sits beside a California rule that grants open release no such discount at all, and beside a White House policy that promotes open release explicitly for its strategic value ^{[5, 6, 7]}. Five prominent 2025 model families settling on plain Apache 2.0 or MIT text sits beside two others — one that has used the same restrictive clause since 2023, and one whose own licence changed between its last two releases — still writing bespoke terms ^{[10, 13, 14, 18, 11, 9, 15, 3]}. Anyone reporting a single confident future for open-weight AI in 2035 is reporting which of these four cells they would bet on, not what the current record shows.

The more useful discipline is the one this article tried to practise: name the axis, ground each pole in a dated figure or a primary licence or statute, and say in advance what observation would mean the world had moved to a different cell. Open-weight AI has already demonstrated real things — a capability gap measured in months rather than years and still narrowing on most readings, a licensing landscape where the majority of prominent 2025 releases chose standard text over custom terms, and a policy landscape where at least one major jurisdiction has already decided compute alone should decide what gets reviewed. Whether any of that closes into the standardised, near-parity, consolidated — or genuinely open and plural — ecosystem its more confident advocates already describe it as becoming is still open, and it will stay open until the comparator, the weighbridge and the sorting table in this article’s own room all report a reading, not just one of them.