Open weights trail the frontier by months. The economics say the gap will not close

The best open-weight model trails the best closed model by about four months. That is Epoch AI’s estimate, published on 29 May 2026, for the period since January 2026: an average gap of 8 points on its Epoch Capabilities Index, which it describes as similar to the distance between GPT-5 and GPT-5.5. Stanford’s 2026 AI Index, using the Arena leaderboard, reports the top closed model 3.3% ahead of the top open model as of March 2026, up from 0.5% in August 2024. The gap is small. It is also no longer shrinking.

That combination is what the economics predict. Below the frontier, open weights have reached parity on nearly everything a production team measures, and they lead on price. At the frontier, the organisations that fund the largest training runs earn their return from exclusive access, and the organisations that give weights away earn their return somewhere else and have no reason to spend the last billion dollars to be first by a quarter. The result is a persistent lag of months, not a convergence to zero.

This piece measures the gap with the primary data, explains how open models closed most of it, argues why the remainder will hold, and sets out what a team building on these models should do about it. The vocabulary matters first, because “open” is doing a lot of work it cannot bear.

Open-weight, open-source and open-data are three different things

An open-weight model publishes its trained parameters. Anyone can download the files, run the model, fine-tune it and, subject to the licence, redistribute the result. That is all the term promises. It says nothing about the training data, the data-processing pipeline, the training code or the evaluation harness, and most open-weight releases publish none of those.

Open source, as the Open Source Initiative defines it, is stricter. The Open Source AI Definition 1.0, released on 28 October 2024, requires four freedoms (use for any purpose without permission, study, modify, share) and three components made available in a form that permits them: “data information”, defined as “sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system”; “the complete source code used to train and run the system”; and the parameters themselves under OSI-approved terms. The definition stops short of requiring the training data itself to be published, a compromise that drew criticism from both directions, but it does require provenance, labelling procedures and sources for public and third-party datasets.

By that standard almost nothing in the 2026 open-weight canon qualifies. Apache 2.0 or MIT on the weights satisfies the parameters clause. The code clause is sometimes met, often partially. The data-information clause is met by a handful of research releases and by NVIDIA’s Nemotron 3 line, which ships training data and recipes under the Linux Foundation’s OpenMDW-1.1 licence, and by very few others. The right term for DeepSeek, Qwen, Kimi, GLM, Llama, Gemma and gpt-oss is open-weight. Moonshot’s own materials, to its credit, use exactly that phrase.

The licence is where the practical differences live, and the licences diverge more than the marketing suggests.

Three things stand out. First, the direction of travel is towards permissive terms: Google moved Gemma from its custom terms to Apache 2.0 in April 2026, and Meta’s return to open releases in August 2026 used Apache 2.0 rather than a Llama-style community licence. Second, the exceptions are the labs that need the weights to pay: Moonshot’s K3 licence adds a clause aimed squarely at hosting providers who resell the model as an API, and Z.ai moved GLM-5.3 off MIT after two MIT generations. Third, a permissive licence on the weights is not the same as a right to the data behind them, which is why the OSI distinction keeps mattering in procurement and in reading a model card properly.

The gap, measured

Three independent measurements exist, and they agree on shape even where they differ on units.

The AI Index numbers show a gap that collapsed through 2024, touched near-zero in August 2024, and then reopened as the closed labs shipped their reasoning generations. The 2026 edition notes that six of the ten highest-scoring models on Arena in March 2026 were proprietary, and that the top six organisations sat within 25 Elo points of each other: Anthropic at 1,503, xAI at 1,495, Google at 1,494, OpenAI at 1,481, Alibaba at 1,449 and DeepSeek at 1,424. The same chapter puts the top US model 2.7% ahead of the top Chinese model, a separate gap that overlaps with the open-closed one because the top open models are Chinese.

Epoch’s method is more useful for the question at hand because it measures time rather than points. Its October 2025 analysis found open-weight models lagging the closed state of the art by about three months on the Epoch Capabilities Index, with a 90% confidence interval of 1.1 to 5.3 months, and noted two historical examples: Llama 3.1-405B was rated on par with Claude 3.5 Sonnet until o1-mini arrived, and DeepSeek-V2 reached approximate parity with GPT-4 about 14 months after GPT-4’s release. The analysis also warned that the gap “likely appears larger than it really is” because the newest open models, gpt-oss-120b and MiniMax-M2 at the time, had too few evaluations to score.

The May 2026 update covers a period with far more open releases and lands at four months and 8 ECI points, with GPT-5.5 Pro at 159.35, Gemini 3.5 Flash at 156.31 and Claude Opus 4.7 at 156.18 on the closed side and Kimi K2.6 at 151.60 and GLM-5 at 146.62 on the open side. Epoch’s own comparison is the telling one: 8 ECI points is roughly the gap between GPT-5 and GPT-5.5. That is a generation, not a rounding error, but it is a generation that arrives every few months.

Artificial Analysis, whose index aggregates ten evaluations including agentic and long-context tasks, put the leading open models at 54 against GPT-5.5’s 60 on 30 April 2026 and observed that all ten of the top open-weight models on its index came from China-based labs. By late September 2026, on version 4.3.2 of the index, its open-weight list is led by Xiaomi’s MiMo-V2.6-Pro at 46, GLM-5.3 at 45 and Kimi K3 at 44; the index was rescaled between the two dates, so the absolute numbers are not comparable, but the ordering shows how quickly the top of the open list turns over. Four different labs held the top open spot between February and September 2026.

One further point from the Artificial Analysis data deserves attention because it explains where the remaining gap sits. On its Omniscience measure, which scores factual knowledge net of hallucination, the open models trail by far more than six points. That matches DeepSeek’s own V4-Pro card, which reports 57.9 on SimpleQA-Verified against 75.6 for Gemini 3.1 Pro even as it reports leads on LiveCodeBench and Codeforces. The open models have caught up on tasks with verifiable answers, where reinforcement learning on checkable problems works. They remain behind on broad world knowledge and calibrated refusal, where scale and proprietary post-training data still tell.

How open models caught up

The catch-up was not one thing. Four mechanisms did most of the work, and all four are now published well enough to be copied.

Mixture-of-experts efficiency

DeepSeek-V3, released in December 2024, activates 37 billion of its 671 billion parameters per token. Its technical report states that the full training run took 2.788 million H800 GPU hours on 14.8 trillion tokens: 2,664 thousand hours of pre-training, 119 thousand of context extension and 5 thousand of post-training. At the paper’s assumed rental price of $2 per GPU hour that is $5.576 million. The report is explicit that the figure “include[s] only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data”, and the number was widely misread as the cost of building DeepSeek. It is nonetheless a real data point: a model that matched GPT-4o on most public benchmarks was trained for single-digit millions of dollars of marginal compute.

Every leading open release since has followed the same architecture. Qwen3.5-397B-A17B activates 17 billion of 397 billion. Kimi K2.6 activates 32 billion of one trillion. DeepSeek-V4-Pro activates 49 billion of 1.6 trillion, and its card claims the new hybrid attention needs 27% of the inference compute of its predecessor. Nemotron 3 Ultra activates 55 billion of 550 billion. MiMo-V2.6-Pro activates 42 billion of 1.02 trillion. Sparse activation cuts both the training bill and the serving bill, and serving cost is what makes an open model economically attractive to host.

Reinforcement learning recipes in the open

DeepSeek-R1, released in January 2025 under MIT, came with a paper whose central claim was that reasoning “can be incentivized purely through RL, without the need for SFT”. Within months the recipe was reproduced at every scale. Qwen3, released on 29 April 2025 under Apache 2.0 with a hybrid thinking mode, cited DeepSeek-R1, o1, o3-mini, Grok-3 and Gemini 2.5 Pro as its comparison set. The pattern since has been that each open lab publishes enough of its post-training method for the next one to build on: Xiaomi’s V2.6 card frames the model around RL scaling rather than supervised fine-tuning, and NVIDIA’s Nemotron 3 Ultra release includes 10 million supervised samples, one million RL tasks and 15 RL environments alongside the weights.

Distillation, in both directions

Distillation is the least discussed and most important mechanism. DeepSeek shipped R1 alongside six distilled models built on Qwen2.5 and Llama 3 bases, with the Qwen-32B distillation reported to outperform o1-mini. Meta’s Llama 4 announcement in April 2025 described the two-trillion-parameter Behemoth as a teacher for Scout and Maverick. OpenAI’s gpt-oss model card states that both models were trained through “large-scale distillation and reinforcement learning”. Meta’s Muse Glimmer, released on 10 August 2026, was trained “on Muse Spark’s outputs using logit distillation”. In each case a frontier-scale model that is not released, or not yet, produces a smaller model that is. Open weights are, increasingly, the distilled shadow of a closed teacher.

The same mechanism runs across company lines, which is the part that worries the closed labs. Zuckerberg’s August 2026 essay argues for removing restrictions on “training data use and distillation” as the way for American open models to compete, which is as close as a frontier CEO has come to saying distillation from rivals is how the open ecosystem keeps pace.

The incentives of Chinese labs

Artificial Analysis’s observation that the top ten open models on its index in April 2026 were all Chinese is not an accident of talent. It follows from who pays. Alibaba announced on 24 February 2025 that it would invest more than RMB 380 billion in AI and cloud infrastructure over three years, more than its total spending on those areas in the previous decade, and its chief executive framed cloud as the clearest AI revenue driver. Qwen is the top of that funnel: the weights are free, the inference runs on Alibaba Cloud, and the proprietary Qwen3.5-Plus sits above the open line as the paid tier. Zhipu became a listed company on the Hong Kong exchange in January 2026 and released GLM-5 under MIT a month later; for a listed model lab, open releases that top the leaderboards are investor relations. DeepSeek was financed by its founder’s quantitative hedge fund rather than venture capital, and has never needed weights to generate revenue to justify a round. Xiaomi sells phones and cars and wants models on its devices.

Why the last few months will not close

Set against those mechanisms is one fact that has not changed: the frontier run is paid for by the expectation of exclusive access, and that expectation is priced in hundreds of billions of dollars of capital expenditure.

The bill

Epoch’s June 2024 analysis of training costs found the amortised hardware and energy cost of the final training run of frontier models growing at 2.4 times per year since 2016, with a cloud-price method giving 2.6 times, and projected that “the largest training runs will cost more than a billion dollars by 2027”. Its breakdown of development cost put hardware at 47 to 67%, research staff at 29 to 49% and energy at 2 to 6%. The final run is only the visible part; the ablations, failed runs and data work sit around it, which is exactly the caveat DeepSeek attached to its $5.576 million.

Meta’s own disclosures give the scale. On 30 July 2025 the company guided to full-year 2025 capital expenditure of $66 to 72 billion and said to expect “another year of similarly significant capital expenditures dollar growth in 2026”. On its second-quarter 2026 release it reported $30.12 billion of capex in the quarter and raised full-year 2026 guidance to $130 to 145 billion. That is the spending of a company that decided, in the same July 2025 letter, that it would be “careful about what we choose to open source”. Its most advanced model, Muse Spark, is served through Meta’s products and listed as proprietary on the Arena leaderboard, where the 1.2 version sat seventh on 25 September 2026. What Meta released in August 2026 was a 30-billion-parameter distillation of it under Apache 2.0, and a promise, repeated in the essay and by its executives, that Muse Spark 1.2 weights would follow. At the time of writing they had not appeared and the Arena still listed the model as proprietary.

The return

A closed lab that spends a billion dollars on a run recovers it through API revenue, subscriptions and enterprise contracts that exist only because nobody else can serve the model. The moment it releases the weights, every hosting provider can serve the same model at marginal cost, and the price collapses to the cost of inference. That is a good outcome for users and a bad one for the lab that paid for training. Epoch’s November 2024 report named this plainly: developers selling access through APIs “have financial incentives to keep their most advanced models proprietary”, and it added a second reason, that releasing highly capable weights is irreversible in a way an API is not.

The organisations that do release weights are those for which the collapse in model price is either harmless or actively helpful. That is why the funders table later in this piece is the real explanation of the gap, and why the lanes below describe two different businesses rather than two different attitudes.

Figure 1: The frontier stays closed because its funders need exclusivity; open weights come from funders who are paid another way.

Each open funder has a reason to reach “as good as the closed model from a few months ago”, because that is the level at which adoption follows. None has a reason to spend the marginal billion to be first by a quarter, because being first by a quarter is only worth a billion dollars if you can charge for exclusive access to it. The gap is the price of that marginal billion, expressed in months.

The moats that are not compute

Two other constraints keep the frontier closed. The first is data. The open models’ weakest results are on knowledge and hallucination, where the closed labs’ advantage comes from proprietary post-training data, human feedback at scale and the interaction logs of consumer products with hundreds of millions of users. Distillation transfers behaviour; it does not transfer the data pipeline that produced it, and the Kimi K3 and GLM-5.3 licence changes show that even open labs now guard the commercial layer.

The second is regulation. Article 51 of the EU AI Act presumes a general-purpose model to have “high impact capabilities”, and therefore systemic risk, when its training compute exceeds 10^25 floating point operations. Article 53(2) exempts providers of models “released under a free and open-source licence” from the technical-documentation obligations in Article 53(1)(a) and (b), but states that “this exception shall not apply to general-purpose AI models with systemic risks”. Every frontier-scale open release crosses the threshold, so the exemption that was written for open models does not reach the ones that matter, and the provider carries the full Chapter V obligations regardless of licence. The obligations have applied since 2 August 2025, which raises the cost of a frontier-scale open release in the EU without raising the return.

Who pays for open weights, and why

Every leading open model of 2026 traces to a funder whose return does not depend on selling the model. The table names them.

Figure 2: Open weights come from organisations whose return does not depend on exclusive access.

The pattern in the right-hand column is the same in every row. The funder releases the model that maximises adoption of the thing it actually sells, and keeps or never builds the model that would compete with that thing. Alibaba keeps Qwen3.5-Plus closed. OpenAI’s gpt-oss-120b, at 117 billion total and 5.1 billion active parameters, runs on a single 80 GB GPU and was released under Apache 2.0 in August 2025; it is a generation behind the paid models and was, by the model card’s own account, distilled. Google’s Gemma 4 moved to Apache 2.0 in April 2026 at up to 31 billion parameters; Gemini did not. Meta released Glimmer at 30 billion parameters, small enough for a laptop, and kept the teacher.

NVIDIA is the clearest case because it has the least reason to hold anything back. Nemotron 3 Ultra, released on 4 June 2026, ships weights, base checkpoints, reward models, quantised variants, training data and recipes under OpenMDW-1.1, a licence the Linux Foundation published on 28 May 2026 to cover “models (including architecture, weight and parameters), code, documentation, or data” in a single instrument. NVIDIA said it will use the same licence across its Cosmos, Isaac GR00T, Ising and Nemotron families. A chip vendor’s ideal world is one in which every model is free and every model needs its hardware, and its release policy is that world in miniature.

The AI Alliance, launched by IBM and Meta in December 2023 and describing itself as a network of companies, universities and government organisations to advance “open, safe, and responsible AI”, is the institutional expression of the same interests: the members are overwhelmingly organisations that sell hardware, hosting, services or research, not model access.

What this means for builders

The measured gap and the funding map together produce a practical rule. For any task the current open frontier can do, the question is which open model, served where, at what cost. For the tail of tasks where four months of capability decide the outcome, pay for the closed model and move on.

Where open weights win

Cost at volume. DeepSeek’s published API prices for V4-Pro, as of September 2026, are $0.66 per million input tokens on a cache miss and $1.98 per million output tokens off-peak, doubling at peak hours; V4-Flash is $0.15 and $0.60. Those are the prices of a lab that does not need the API to fund training, and every hosting provider serving the same weights competes down to a similar level. The 2025 AI Index recorded a 280-fold fall in the inference cost of a GPT-3.5-level system between November 2022 and October 2024, and open weights are the mechanism that keeps that curve falling: a closed model’s price floor is set by its owner, an open model’s by the cheapest competent host.

Data residency and privacy. Weights you hold can run inside a network boundary, in a jurisdiction of your choosing, without a data-processing agreement with a model provider. Nothing about a closed API replaces that, whatever its contract says.

Latency and control. Self-hosting sets the batching, quantisation and caching policy, and removes the provider’s queue from the critical path. gpt-oss-20b runs in 16 GB of memory; Glimmer fits a 24 or 32 GB consumer GPU after quantisation. For interactive products where the round trip matters more than the last few points of capability, a small open model close to the user beats a large closed one across the internet.

Fine-tuning and ownership. A fine-tuned open model is an asset the team owns and can move between hosts. A fine-tuned closed model is a feature of one provider.

Where they do not

The knowledge tail. The SimpleQA-Verified and Omniscience gaps are the concrete version of “reliability at the tail”. A task that depends on obscure facts, calibrated abstention or refusal behaviour tuned on proprietary data is a task where the closed models still earn their price.

Long-horizon agents on the newest capability. Epoch’s 8-point gap is the difference between a model generation. If the product is defined by what the newest generation can do, the open model will reach it in a few months and the product cannot wait.

Operational load. Self-hosting means owning capacity planning, GPU procurement or reservation, serving-stack upgrades and the monitoring that comes with them. At low or bursty volume, a hosted open-model API is nearly always cheaper than owned hardware, and the arithmetic should be done per workload rather than as a policy.

Licence risk. Apache 2.0 and MIT are safe. The Llama 4 community licence, the Kimi K3 model-as-a-service clause and the GLM-5.3 custom licence are each fine for most users and a problem for some, particularly anyone reselling inference. Read the file, not the press release.

A team that routes through an LLM gateway can make this a configuration decision rather than an architectural one: open weights for the bulk of traffic, a closed model for the flagged minority, and a switch when the next open release lands. Given that the top open spot changed hands four times between February and September 2026, the switch will be used.

Figure 3: Open versus closed is rarely the decision. Where to serve which model, for which slice of traffic, is.

The counter-argument: commoditise the layer

The strongest case against this analysis is Linux. In the 1990s the frontier of operating systems was closed, expensive and funded by exclusive licensing, and it was overtaken not by an open project reaching parity on the day of each release but by an open project becoming the layer everyone built on, until the closed frontier stopped mattering for most uses. On this view the four-month gap is irrelevant. What matters is that five or more independent open families now sit within a generation of the frontier, that every hosting provider can serve them, and that the price of model capability is being set by the cheapest open host rather than by the owner of the best model. The closed labs’ capex is then a bet on being the exclusive supplier of a product whose price is falling towards its open substitute.

There is evidence for this. Meta’s Zuckerberg wrote in August 2026 that restricting access to foreign open models would “reduce the quality of AI accessible to them, and centralize AI rather than putting more power in people’s hands”, and that “falling behind in AI overall would almost certainly be a larger and longer term national security issue” than any single risk of releasing weights. That is a company with $130 to 145 billion of 2026 capex arguing that the layer should be commoditised, because Meta does not sell the layer.

Where the analogy breaks is the cost curve. Linux caught up with a product whose marginal cost of development was flat; the frontier here is a product whose training cost grows 2.4 times a year and whose next run will cost more than a billion dollars. Commoditising the layer requires someone to keep paying that bill for a model they then give away, and the funders table shows that nobody does: they give away the distillation and keep the teacher. Linux won because the frontier stopped moving. This frontier has not.

The honest version of the counter-argument is narrower and probably right: for the majority of workloads, the open layer is already the commodity, and the closed labs’ pricing power is confined to the moving edge. That is not the same as the gap closing. It is the gap becoming irrelevant to most buyers while remaining decisive for a few.

Our judgement

The gap between the best open-weight and the best closed model is about four months by Epoch’s measure and about 3% by Stanford’s, and it has been roughly that for a year. We think it stays there. The mechanisms that closed the gap from a year to a quarter, sparse architectures, published RL recipes and distillation, are now standard and cannot close it further, because the last quarter is bought with the marginal billion dollars that only an exclusive-access business model repays. The organisations that release weights are, without exception, paid another way, and they release the model that maximises what they are paid for.

What to do with that:

1. Default to open weights for anything the current open frontier handles, and re-run the comparison quarterly. The top open model changed four times in the first nine months of 2026.
2. Reserve closed models for the knowledge tail and the newest capability, and measure which of your tasks are actually there. It is usually a smaller share than the team assumes.
3. Choose the licence before the benchmark. Apache 2.0 and MIT for anything you might resell; read the Llama, Kimi K3 and GLM-5.3 terms before building on them.
4. Treat self-hosting as a volume decision, not an ideology. Steady, high, latency-sensitive or residency-constrained traffic justifies owned GPUs; the rest is cheaper on a hosted open-model API.
5. Watch three signals. Whether Meta ships Muse Spark 1.2 weights as promised, whether Epoch’s four-month figure moves in either direction, and whether the next licence changes at the open labs tighten further. Each will say more about the gap than any single benchmark.

The frontier will stay closed for the tasks that need it. Everything else is increasingly running on weights anyone can download, priced by whoever hosts them most cheaply. That is the more consequential trend, and it does not require the gap to close.